CN116630812B

CN116630812B - Water body feature detection method and system based on visible light image analysis

Info

Publication number: CN116630812B
Application number: CN202310900019.0A
Authority: CN
Inventors: 李春林; 严鹏; 郑能建; 王佳; 龚雪
Original assignee: Sichuan Development Environmental Science And Technology Research Institute Co ltd
Current assignee: Sichuan Development Environmental Science And Technology Research Institute Co ltd
Priority date: 2023-07-21
Filing date: 2023-07-21
Publication date: 2023-09-26
Anticipated expiration: 2043-07-21
Also published as: CN116630812A

Abstract

The embodiment of the application provides a water body feature detection method and a system based on visible light image analysis, which are used for carrying out training image evaluation on a plurality of sample water body acquisition image sequences based on a training image evaluation network after training is completed so as to screen out a sample water body acquisition image sequence for carrying out water body pollution feature learning, training a water body pollution feature learning model based on the sample water body acquisition image sequence so as to obtain a training completed water body pollution detection model, carrying out water body pollution detection on an input water body acquisition image of a target water body area based on the training completed water body pollution detection model so as to obtain water body pollution detection features of the input water body acquisition image, and carrying out visual information prompt on a management service terminal corresponding to the target water body area based on the water body pollution detection features of the input water body acquisition image, thereby carrying out model training after the training image evaluation, carrying out water body pollution detection and improving the effectiveness of water body pollution detection.

Description

Water body feature detection method and system based on visible light image analysis

Technical Field

The embodiment of the application relates to the technical field of image analysis, in particular to a water body characteristic detection method and system based on visible light image analysis.

Background

The water pollution refers to the phenomenon that when the pollutant entering the water exceeds the environment capacity of the water or the self-cleaning capacity of the water, the water quality is deteriorated, so that the original value and the function of the water are destroyed. In the related art, in order to perform water pollution detection, water is usually investigated in each area by on-site inspection personnel, and water pollution detection is performed by a manual sampling detection scheme, which is time-consuming and labor-consuming and requires a great deal of time for a great deal of staff to participate. Therefore, in the related art, a scheme of performing pollution characteristic analysis of a water body image based on an image AI technology is proposed, however, the related art does not perform validity assessment of a training image, and therefore, the generated neural network model still has difficulty in guaranteeing validity of water body pollution detection.

Disclosure of Invention

In order to at least overcome the defects in the prior art, the embodiment of the application aims to provide a water body characteristic detection method and system based on visible light image analysis.

According to an aspect of the embodiment of the application, there is provided a method for detecting a water feature based on visible light image analysis, including:

performing training image evaluation on the plurality of sample water body acquisition image sequences based on the training image evaluation network after training is completed so as to screen out sample water body acquisition image sequences for performing water body pollution characteristic learning;

Training the water pollution characteristic learning model based on the sample water body acquisition image sequence to obtain a trained water pollution detection model;

performing water pollution detection on an input water body acquisition image of a target water body area based on the trained water body pollution detection model to obtain water body pollution detection characteristics of the input water body acquisition image;

and carrying out visual information prompt on the management service terminal corresponding to the target water body area based on the water body pollution detection characteristics of the input water body acquisition image.

According to one aspect of the embodiment of the present application, there is provided a visible light image analysis-based water feature detection system, including a processor and a machine-readable storage medium having stored therein machine-executable instructions loaded and executed by the processor to implement the visible light image analysis-based water feature detection method in any one of the foregoing possible embodiments.

According to an aspect of embodiments of the present application, there is provided a computer program product or computer program comprising computer instructions stored in a computer readable storage medium. The computer instructions are read from a computer-readable storage medium by a processor of a computer device, and executed by the processor, cause the computer device to perform the methods provided in the various alternative implementations of the above aspects.

The embodiment of the application has the beneficial effects that:

the training image evaluation network is used for carrying out training image evaluation on a plurality of sample water body acquisition image sequences based on the training image evaluation network to screen out sample water body acquisition image sequences for carrying out water body pollution characteristic learning, the water body pollution characteristic learning model is trained based on the sample water body acquisition image sequences to obtain a training-completed water body pollution detection model, the input water body acquisition image of the target water body area is subjected to water body pollution detection based on the training-completed water body pollution detection model to obtain water body pollution detection characteristics of the input water body acquisition image, and the management service terminal corresponding to the target water body area is subjected to visual information prompt based on the water body pollution detection characteristics of the input water body acquisition image, so that model training is carried out after the training image evaluation, thereby carrying out water body pollution detection and improving the effectiveness of water body pollution detection.

Drawings

Fig. 1 is a schematic flow chart of a method for detecting water body characteristics based on visible light image analysis according to an embodiment of the present application;

fig. 2 is a schematic block diagram of a visible light image analysis-based water feature detection system for implementing the visible light image analysis-based water feature detection method according to an embodiment of the present application.

Detailed Description

Fig. 1 is a schematic flow chart of a method for detecting water body characteristics based on visible light image analysis according to an embodiment of the present application, and the method for detecting water body characteristics based on visible light image analysis is described in detail below.

Step S100, training image evaluation is carried out on a plurality of sample water body acquisition image sequences based on the training image evaluation network after training is completed, so as to screen out sample water body acquisition image sequences for carrying out water body pollution characteristic learning.

In this embodiment, the training image evaluation network may be configured to evaluate subsequent training values of the plurality of sample water body acquisition image sequences, where a higher training value indicates a greater training reliability of the sample water body acquisition images in the corresponding sample water body acquisition image sequence. Through training image evaluation, a sample water body acquisition image sequence for carrying out water body pollution characteristic learning can be screened out.

Step 200, training the water pollution characteristic learning model based on the sample water body acquisition image sequence to obtain a trained water pollution detection model.

In this embodiment, the training value is used to evaluate the image sequence collected by the screened sample water body, so as to train the learning model of the water pollution characteristics, and a water pollution detection model with the water pollution detection capability can be obtained.

And step S300, carrying out water pollution detection on the input water body acquisition image of the target water body area based on the trained water body pollution detection model, and obtaining the water body pollution detection characteristics of the input water body acquisition image.

And step S400, carrying out visual information prompt on the management service terminal corresponding to the target water body area based on the water body pollution detection characteristics of the input water body acquisition image.

Based on the steps, training image evaluation is carried out on a plurality of sample water body acquisition image sequences based on a training image evaluation network after training is completed so as to screen out a sample water body acquisition image sequence for carrying out water body pollution characteristic learning, a water body pollution characteristic learning model is trained based on the sample water body acquisition image sequence so as to obtain a training completed water body pollution detection model, water body pollution detection is carried out on an input water body acquisition image of a target water body area based on the training completed water body pollution detection model so as to obtain water body pollution detection characteristics of the input water body acquisition image, and visual information prompt is carried out on a management service terminal corresponding to the target water body area based on the water body pollution detection characteristics of the input water body acquisition image, so that model training is carried out after the training image evaluation, thereby carrying out water body pollution detection, and improving the effectiveness of the water body pollution detection.

In an exemplary design concept, for step S100, it may be implemented by the following steps.

Step S101, acquiring a plurality of sample water body acquisition image sequences.

The training value network units in the training image evaluation network are used for updating parameters of the training value network units in the training image evaluation network, and the number of the training value network units in the training image evaluation network is the same as the number of the sample water body acquisition image sequences. Wherein any one of the plurality of example water body acquisition image sequences includes an active example water body acquisition image, a passive example water body acquisition image, and an unsupervised example water body acquisition image; the active example water body acquisition image is an image with high training value, namely, an example image corresponding to a training effective image source comprises a source acquisition vector of the training effective image source and active label training value; the non-supervision example image is an example image with unknown training value, namely an example image of an image source with unknown training effectiveness, and comprises source acquisition vectors of the image source with unknown training effectiveness; the negative example water body acquisition image is an image with a non-high training value, namely, the example image corresponding to the non-training effective image source comprises a source acquisition vector of the non-training effective image source and a negative label training value. The negative example water body acquisition image is obtained by labeling the labeling training value of the target water body image, which is not matched with the labeling training value in the observation description in the first observation example water body acquisition image, and the target water body image comprises a source acquisition vector of a non-training effective image source. Thus, in other words, the negative example water body acquisition image includes a target water body image in the first observed example water body acquisition image for which the observed description does not match the annotation training value. The observation description is description information generated by training image evaluation on the target water body image according to a training image evaluation network, and the labeling training value is training value generated by training value calibration on the target water body image; in other words, the observation description indicates that the target water body image is an image with high training value, i.e., an example image corresponding to a training effective image source, and the labeling training value of the target water body image indicates that the target water body image is an image with non-high training value, i.e., an example image corresponding to a non-training effective image source.

In some exemplary design considerations, acquiring a plurality of example water acquisition image sequences may include: acquiring a reference water body acquisition image sequence, wherein the reference water body acquisition image sequence comprises a plurality of supervised active sample images and a plurality of unsupervised sample images, any one of the plurality of supervised active sample images comprises a source acquisition vector of a training effective image source, and any one of the plurality of unsupervised sample images comprises a source acquisition vector of an image source with unknown training effectiveness; constructing a plurality of sample water body acquisition image sequences, wherein positive sample water body acquisition images in any one sample water body acquisition image sequence are randomly extracted from a plurality of supervised positive sample images in a reference water body acquisition image sequence, non-supervised sample water body acquisition images in any one sample water body acquisition image sequence are randomly extracted from a plurality of non-supervised sample images in the reference water body acquisition image sequence, and negative sample water body acquisition images in any one sample water body acquisition image sequence are acquired after training image evaluation of a first observation sample water body acquisition image through a training image evaluation network.

In some exemplary design concepts, the supervised active exemplar images included in the reference water acquisition image sequence are high training value images including source acquisition vectors of training efficient image sources and active annotation training values, the unsupervised exemplar images included in the reference water acquisition image sequence are exemplar images of unknown training values including source acquisition vectors of training efficient image sources and not including annotation training values for representing the training values of the exemplar images, i.e., neither active nor passive annotation training values. The positive annotation training value can be manually annotated based on priori knowledge.

In some exemplary design considerations, the positive example water body acquisition image in any one of the constructed plurality of example water body acquisition image sequences is randomly extracted from the plurality of supervised positive example images in the reference water body acquisition image sequence, and the non-supervised example water body acquisition image in any one of the example water body acquisition image sequences is randomly extracted from the plurality of non-supervised example images in the reference water body acquisition image sequence. Taking an example water body collection image sequence as an example, a set number of supervised active example images may be randomly extracted from a plurality of supervised active example images in a reference water body collection image sequence as active example water body collection images, and a set number of non-supervised example images may be randomly extracted from a plurality of non-supervised example images in the reference water body collection image sequence as non-supervised example water body collection images. The number of supervised active and non-supervised example images may be specifically set based on specific requirements, e.g., the same number of supervised active and non-supervised example images may be extracted. As another example, 70% of the supervised active exemplar images may be randomly extracted from the plurality of supervised active exemplar images in the sequence of reference water acquisition images as the active exemplar water acquisition images.

Since the example water body acquisition image sequence is randomly extracted according to the supervised active and the unsupervised example images in the reference water body acquisition image sequence, the active and the unsupervised example water body acquisition images included in the different example water body acquisition image sequences thus obtained are generally different from each other as the unsupervised example water body acquisition image 1. According to the random extraction mode from the same reference water body acquisition image sequence, the condition that the number of images in the reference water body acquisition image sequence is small can be fully dealt with, and the example images in the reference water body acquisition image sequence can be fully utilized. For example, if the sequence of reference water acquisition images is { F1, F2, F3, F4, F5, G1, G2, G3, G4, G5, G6, G7, G8, G9, G10}, F1, F2, F3, F4, F5 are all supervised positive example images, G1, G2, G3, G4, G5, G6, G7, G8, G9, G10 are all non-supervised example images; if 3 sample water body acquisition image sequences need to be constructed, each sample water body acquisition image sequence needs to include 4 active sample water body acquisition images and 4 unsupervised sample water body acquisition images, one condition of the active sample water body acquisition images and the unsupervised sample water body acquisition images in the obtained 3 sample water body acquisition image sequences may be: the active and unsupervised example water body acquisition images in the 1 st example water body acquisition image sequence are { F1, F2, F3, F4, G1, G2, G3, G4}, the active and unsupervised example water body acquisition images in the 2 nd example water body acquisition image sequence are { F2, F3, F4, F5, G2, G3, G4, G5}, and the active and unsupervised example water body acquisition images in the 3 rd example water body acquisition image sequence are { F1, F3, F4, F5, G6, G7, G8}.

In some exemplary design concepts, the negative example water body acquisition image in any one of the example water body acquisition image sequences is acquired after a training image evaluation of the first observed example water body acquisition image by a training image evaluation network. Illustratively, the negative example water body acquisition image is obtained by corresponding the annotation training value to a target water body image in which the observation description does not match the annotation training value in the first observation example water body acquisition image. In other words, the target water body image is a first observation sample water body acquisition image of the training image evaluation network evaluation error, specifically: when the training image evaluation network performs training evaluation on the first observation sample water body acquisition image, the first observation sample water body acquisition image with non-high training value is judged to be the first observation sample water body acquisition image with non-high training value corresponding to the first observation sample water body acquisition image with high training value. The training value of the annotation carried by the target water body image with the incorrect evaluation of the training image evaluation network can be calibrated by a user, namely, the user can screen the first observation sample water body acquisition image of which the observation description of the first observation sample water body acquisition image indicates an image with high training value, the first observation sample water body acquisition image with the incorrect evaluation of the training image evaluation network is generated and used as the target water body image, and the target water body image is annotated with the negative annotation training value, so that the negative sample water body acquisition image is generated. Wherein each of the sequence of example water body acquisition images includes passive example water body acquisition images that are comprised of the same target water body image.

Step S102, parameter updating is performed on a plurality of training value network units in the training image evaluation network according to a plurality of sample water body acquisition image sequences, a target threshold training value of each training value network unit in the plurality of training value network units is generated, the target threshold training value of each training value network unit in the plurality of training value network units is determined as the threshold training value of each updated training value network unit, and a plurality of updated training value network units are generated.

Wherein, the different example water body acquisition image sequences correspond to different training value network units.

In some exemplary design considerations, the training value network element may be a support vector machine. The training image evaluation network comprises a certain number of training value network units, and the number of the training value network units in the training image evaluation network can be determined according to actual requirements.

In some exemplary design ideas, parameter updating is performed on a plurality of training value network elements in a training image evaluation network according to a plurality of sample water body acquisition image sequences, and generating a target gate training value for each training value network element in the plurality of training value network elements may include: taking the non-supervision example water body acquisition image as a negative example water body acquisition image, carrying out parameter updating on a plurality of training value network units in a training image evaluation network according to a plurality of example water body acquisition image sequences, and calculating network performance indexes of each training value network unit under different threshold training values; and determining the threshold training value corresponding to the maximum network performance index in the network performance indexes of each training value network unit as the target threshold training value of each training value network unit in the plurality of training value network units. The network performance index can reflect the performance of each training value network unit under different threshold training values, and is a harmonic average value for measuring the precision rate and recall rate of the training value network units.

Taking a training value network unit as an example, taking a source acquisition vector of a training effective image source included in an active sample water body acquisition image in a sample water body acquisition image sequence, a source acquisition vector of a non-training effective image source included in a passive sample water body acquisition image in the sample water body acquisition image sequence, and a source acquisition vector of an image source with unknown training effectiveness included in an unsupervised sample water body acquisition image in the sample water body acquisition image sequence as inputs of the training value network unit, through which training value network unit, the training value of each active sample water body acquisition image, the training value of each passive sample water body acquisition image, and the training value of each unsupervised sample water body acquisition image can be obtained; the training value network unit may then determine training value observations for each positive example water body acquisition image, training value observations for each negative example water body acquisition image, and training value observations for each unsupervised example water body acquisition image based on the threshold training values of the training value network unit. Illustratively, the training value network unit determines a training value observation result of the active example water body acquisition image with the training value greater than the threshold training value of the training value network unit in the training value of each active example water body acquisition image as an active result, that is, the training value network unit outputs the active example water body acquisition image as an image with high training value, that is, the training value network unit outputs a training effective image source corresponding to the active example water body acquisition image as a training effective image source; the training value network unit determines the training value observation result of the active example water body acquisition image with the training value not larger than the threshold training value of the training value network unit in the training value of each active example water body acquisition image as a negative result, namely the training value network unit outputs the active example water body acquisition image as an image with a non-high training value, namely the training value network unit outputs a training effective image source corresponding to the active example water body acquisition image as a non-training effective image source; the training value network unit determines a training value observation result of the passive example water body acquisition image with the training value larger than the threshold training value of the training value network unit in the training value of each passive example water body acquisition image as a positive result, namely the training value network unit outputs the passive example water body acquisition image as an image with high training value, namely the training value network unit outputs a non-training effective image source corresponding to the passive example water body acquisition image as a training effective image source; the training value network unit determines a training value observation result of the passive example water body acquisition image with the training value not larger than the threshold training value of the training value network unit in the training value of each passive example water body acquisition image as a passive result, namely the training value network unit outputs the passive example water body acquisition image as an image with a non-high training value, namely the training value network unit outputs a non-training effective image source corresponding to the passive example water body acquisition image as a non-training effective image source; the training value network unit determines the training value observation result of the non-supervised example water body acquisition image with the training value larger than the threshold training value of the training value network unit in the training value of each non-supervised example water body acquisition image as a positive result, namely the training value network unit outputs the non-supervised example water body acquisition image as an image with high training value, namely the training value network unit outputs an image source with unknown training effectiveness corresponding to the non-supervised example water body acquisition image as a training effective image source; the training value network unit determines the training value observation result of the non-supervised example water body acquisition image with the training value not larger than the threshold training value of the training value network unit in the training value of each non-supervised example water body acquisition image as a negative result, namely the training value network unit outputs the non-supervised example water body acquisition image as an image with a non-high training value, namely the training value network unit outputs an image source with unknown training effectiveness corresponding to the non-supervised example water body acquisition image as a non-training effective image source. Therefore, the network performance index of the training value network element under different threshold training values can be calculated; and determining a threshold training value corresponding to the maximum network performance index in the network performance indexes of the training value network unit as a target threshold training value of the training value network unit, and then determining the target threshold training value of the training value network unit as an updated threshold training value of the training value network unit. For example, the non-supervised example water body acquisition image can be used as a negative example water body acquisition image, and network performance indexes under different threshold training values can be calculated.

For example, if the sample water body acquisition image sequence is { F1, F2, F3, F4, F5, H1, H2, H3, K1, K2}, where F1, F2, F3, F4, F5 are all positive sample water body acquisition images, H1, H2, H3 are all negative sample water body acquisition images, K1, K2 are all non-supervised sample water body acquisition images, and the non-supervised sample water body acquisition images are taken as negative sample water body acquisition images, the annotation training value corresponding to the sample water body acquisition image sequence is {1,1,1,1,1,0,0,0,0,0},1 represents the positive annotation training value, and 0 represents the negative annotation training value. If the training value network unit is used for generating the training value of each active example water body acquisition image and the training value of each passive example water body acquisition image in the example water body acquisition image sequence to be {0.75,0.65,0.8,0.9,0.9,0.1,0.55,0.4,0.65,0.9}; then, when the threshold training value of the training value network element is 0.5, the training value observation corresponding to the sample water body acquisition image sequence (i.e., the training value observation of the active sample water body acquisition image and the training value observation of the passive sample water body acquisition image) may be represented as {1,1,1,1,1,0,1,0,1,1}; when the threshold training value of the training value network unit is 0.6, the training value observation result corresponding to the sample water body acquisition image sequence can be represented as {1,1,1,1,1,0,0,0,1,1}; when the threshold training value of the training value network element is 0.7, the training value observation corresponding to the example water body acquisition image sequence may be represented as {1,0,1,1,1,0,0,0,0,1}.

When the threshold training value of the training value network element is 0.5, the network performance index of the training value network element is 0.77, when the threshold training value of the training value network element is 0.6, the network performance index of the training value network element is 0.83, when the threshold training value of the training value network element is 0.7, the network performance index of the training value network element is 0.8, the threshold training value corresponding to the maximum network performance index of the training value network element of 0.83 is determined to be the target threshold training value of the training value network element, and then the target threshold training value of the training value network element is determined to be the updated threshold training value of the training value network element. The target gate training value of the training value network unit can also be set in a self-defined manner.

Step S103, updating the threshold training value of the training image evaluation network according to the threshold training values of the plurality of updated training value network units, and generating an optimized training image evaluation network.

The threshold training value of the optimized training image evaluation network is determined according to the threshold training values of the plurality of updated training value network units.

In some exemplary design considerations, updating the threshold training value of the training image evaluation network according to the threshold training values of the plurality of updated training value network elements may include: processing the threshold training values of the plurality of updated training value network units to generate the threshold training value of the optimized training image evaluation network; and generating an optimized training image evaluation network according to the plurality of updated training value network units and the threshold training value of the optimized training image evaluation network.

In some exemplary design considerations, processing the threshold training values of the plurality of updated training value network elements to generate the optimized training image evaluation network threshold training value may include: processing the threshold training values of the plurality of updated training value network units to generate a reference threshold training value; determining the threshold training value of the optimized training image evaluation network, wherein the number of the threshold training values of the optimized training image evaluation network is a plurality of, the threshold training value of the optimized training image evaluation network at least comprises a reference threshold training value, the plurality of the threshold training values of the optimized training image evaluation network form a plurality of threshold value intervals, and different threshold value intervals correspond to different observation descriptions. And then determining a second observation sample water body acquisition image which can be used for carrying out water body pollution characteristic learning based on the training evaluation score, and constructing a sample water body acquisition image sequence.

The threshold training values of the plurality of updated training value network units may be averaged to generate a reference threshold training value. For example, if there are 5 updated training value network elements, the threshold training value of the 1 st updated training value network element is 0.6, the threshold training value of the 2 nd updated training value network element is 0.7, the threshold training value of the 3 rd updated training value network element is 0.6, the threshold training value of the 4 th updated training value network element is 0.6, and the threshold training value of the 5 th updated training value network element is 0.7, the reference threshold training value may be 0.64 ((0.6+0.7+0.6+0.6+0.7)/(5).

In some exemplary design ideas, the threshold training value of the optimized training image evaluation network may be determined according to the reference threshold training value and the preset threshold training value. The preset threshold training value is set in a self-defining mode based on actual requirements.

In some exemplary design ideas, if the training value network element is a support vector machine, acquiring a reference water body acquisition image sequence comprising a plurality of supervised active paradigms images and a plurality of unsupervised paradigms images; constructing a plurality of sample water body acquisition image sequences, wherein each sample water body acquisition image sequence comprises active sample water body acquisition images, passive sample water body acquisition images and non-supervision sample water body acquisition images, the active sample water body acquisition images in any one sample water body acquisition image sequence are randomly extracted from a plurality of supervised active sample images in a reference water body acquisition image sequence, the non-supervision sample water body acquisition images in any one sample water body acquisition image sequence are randomly extracted from a plurality of non-supervision sample images in the reference water body acquisition image sequence, and the passive sample water body acquisition images in any one sample water body acquisition image sequence are acquired after training image evaluation on a first observation sample water body acquisition image through a training image evaluation network; parameter updating is carried out on different support vector machines in a training image evaluation network according to different sample water body acquisition image sequences, and a plurality of updated support vector machines and threshold training values of each updated support vector machine are generated; generating the threshold training value of the optimized training image evaluation network according to the threshold training value of each updated support vector machine; and generating an optimized training image evaluation network according to the updated support vector machines and the threshold training value of the optimized training image evaluation network.

Step S104, training image evaluation is carried out on the second observation sample water body acquisition image through the optimized training image evaluation network, and an observation description of the second observation sample water body acquisition image is generated.

Wherein, the second observation example water body acquisition image can be a source acquisition vector of any one content producer; the observation description of the water body acquisition image of the second observation example is obtained according to the threshold training value of the optimized training image evaluation network; the observed description of the second observed example water body acquisition image is used for indicating the training value of the content producer corresponding to the second observed example water body acquisition image.

Step S105, determining a target training evaluation score indicated by the observation description of the second observation sample water body acquisition image, wherein the target training evaluation score is one training evaluation score in a plurality of training evaluation scores, and different training evaluation scores in the plurality of training evaluation scores correspond to different threshold value intervals in a plurality of threshold value intervals;

step S106, determining a second observation sample water body acquisition image which can be used for carrying out water body pollution characteristic learning based on the training evaluation score, and constructing a sample water body acquisition image sequence.

In some exemplary design considerations, training value observations are performed on the second observation example water body acquisition image by each updated training value network element in the optimized training image evaluation network, generating a plurality of training value observations of the second observation example water body acquisition image; generating a global training value observation result of the second observation example water body acquisition image according to the plurality of training value observations of the second observation example water body acquisition image; and determining the observation description corresponding to the target gate numerical value interval to which the global training value observation result of the second observation example water body acquisition image belongs as the observation description of the second observation example water body acquisition image. Wherein one of the plurality of training value observations of the second observation example water body acquisition image corresponds to an updated training value network element, the plurality of training value observations of the second observation example water body acquisition image comprising one or more of: positive results and negative results; the global training value observation result of the second observation example water body acquisition image is used for reflecting the ratio of the positive result in the multiple training value observations of the second observation example water body acquisition image to the multiple training value observations of the second observation example water body acquisition image; the target threshold value interval is one threshold value interval of a plurality of threshold value intervals determined according to the threshold training value of the optimized training image evaluation network.

In some exemplary design considerations, when training value observations of the second observation example water body acquisition image are made by each updated training value network element in the optimized training image evaluation network, generating a plurality of training value observations of the second observation example water body acquisition image, one of the plurality of training value observations of the second observation example water body acquisition image corresponding to one updated training value network element, the plurality of training value observations of the second observation example water body acquisition image including one or more of: positive results and negative results. If one training value observation result of the second observation sample water body acquisition image is a positive result, the updated training value network unit corresponding to the positive result is characterized to output the second observation sample water body acquisition image as an image with high training value, namely the updated training value network unit outputs a content producer corresponding to the second observation sample water body acquisition image as a training effective image source. If the other training value observation result of the second observation example water body acquisition image is a negative result, the updated training value network unit corresponding to the negative result is characterized in that the second observation example water body acquisition image is output as an image with a non-high training value, namely the updated training value network unit outputs a content producer corresponding to the second observation example water body acquisition image as a non-training effective image source.

In some exemplary design considerations, the global training value observation of the second observation example water body acquisition image is used to reflect a ratio of positive ones of the plurality of training value observations of the second observation example water body acquisition image to the plurality of training value observations of the second observation example water body acquisition image when the global training value observation of the second observation example water body acquisition image is generated from the plurality of training value observations of the second observation example water body acquisition image. For example, if there are 5 updated training value network elements, training value observations of the second observation example water body acquisition image obtained by the 1 st to 5 th updated training value network elements are respectively: positive results, negative results, positive results; the global training value observation of the second observation example water body acquisition image is 0.8.

In some exemplary design ideas, when the observation description corresponding to the target gate value interval to which the global training value observation result of the second observation example water body acquisition image belongs is determined as the observation description of the second observation example water body acquisition image, the target gate value interval is one of a plurality of gate value intervals determined according to the threshold training value of the optimized training image evaluation network.

In some exemplary design ideas, parameter updating can be performed on a plurality of training value network units in the training image evaluation network according to the acquired plurality of sample water body acquisition image sequences to generate a plurality of updated training value network units; updating the threshold training values of the training image evaluation network according to the threshold training values of the plurality of updated training value network units, and generating an optimized training image evaluation network; wherein, any one of the plurality of example water body acquisition image sequences includes an active example water body acquisition image, a passive example water body acquisition image, and an unsupervised example water body acquisition image, the passive example water body acquisition image including: the method comprises the steps that an observation description and a labeling training value are not matched in a first observation sample water body acquisition image, the observation description is description information generated by training image evaluation on the target water body image according to a training image evaluation network, and the labeling training value is a training value generated by training value calibration on the target water body image. After the optimized training image evaluation network is obtained, training image evaluation can be performed on the second observation sample water body acquisition image through the optimized training image evaluation network, and an observation description of the second observation sample water body acquisition image is generated, wherein the observation description of the second observation sample water body acquisition image is obtained according to the threshold training value of the optimized training image evaluation network. Because the negative example water body acquisition image comprises a target water body image with unmatched observation description and labeling training value in the first observation example water body acquisition image, the observation description is description information generated by training image evaluation on the target water body image according to a training image evaluation network, the labeling training value is training value generated by training value calibration on the target water body image, in other words, the target water body image is the first observation example water body acquisition image with wrong evaluation by the training image evaluation network, and the method specifically comprises the following steps: when the training image evaluation network performs training evaluation on the first observation sample water body acquisition image, judging the first observation sample water body acquisition image with a non-high training value as a first observation sample water body acquisition image with a non-high training value corresponding to the first observation sample water body acquisition image with a high training value; the negative example water body acquisition image is obtained by performing training value calibration and labeling on the target water body image, specifically labeled as negative labeling training value, and represents that the target water body image is an image with a non-high training value. And optimizing and updating the training image evaluation network according to the first observation sample water body acquisition image of the evaluation error of the training image evaluation network, so that the reliability of the evaluation of the high training value image by the optimized training image evaluation network is higher.

Further embodiments are provided below, which may include the steps of:

in step S201, a plurality of sample water acquisition image sequences are acquired.

Step S202, parameter updating is performed on a plurality of training value network units in the training image evaluation network according to a plurality of sample water body acquisition image sequences, a target threshold training value of each training value network unit in the plurality of training value network units is generated, the target threshold training value of each training value network unit in the plurality of training value network units is determined as the threshold training value of each updated training value network unit, and a plurality of updated training value network units are generated.

Step S203, updating the threshold training value of the training image evaluation network according to the threshold training values of the plurality of updated training value network units, and generating an optimized training image evaluation network.

Step S204, training image evaluation is carried out on the second observation sample water body acquisition image through the optimized training image evaluation network, and an observation description of the second observation sample water body acquisition image is generated.

The steps S201 to S204 may be referred to the corresponding descriptions of the steps S101 to S104.

Step S205, updating the target water body image according to the observation description of the second observation sample water body acquisition image and the labeling training value carried by the second observation sample water body acquisition image, and generating an adjusted target water body image.

The observation description of the adjusted target water body image is not matched with the labeling training value of the adjusted target water body image, and the labeling training value carried by the second observation sample water body acquisition image can be set by a user based on priori labeling experience. The adjusted target water body image is a second observation sample water body acquisition image of an evaluation error of an optimized training image evaluation network, and specifically comprises the following steps: when the optimized training image evaluation network performs training evaluation on the second observation sample water body acquisition image, judging the second observation sample water body acquisition image with a non-high training value as a second observation sample water body acquisition image with a non-high training value corresponding to the second observation sample water body acquisition image with a non-high training value; the annotation training value of the adjusted target water body image is a negative annotation training value.

Step S206, updating the sample water body acquisition image sequence.

The negative example water body collection images included in the adjusted example water body collection image sequence are obtained according to the adjusted target water body images and the labeling training values of the adjusted target water body images, in other words, the negative example water body collection images included in the adjusted example water body collection image sequence include the adjusted target water body images and the negative labeling training values. Optionally, the positive example water body acquisition image in the adjusted example water body acquisition image sequence may be re-randomly extracted from the plurality of supervised positive example images in the reference water body acquisition image sequence, and the non-supervised example water body acquisition image in the adjusted example water body acquisition image sequence may be re-randomly extracted from the plurality of non-supervised example images in the reference water body acquisition image sequence.

Step S207, the adjusted sample water body acquisition image sequence is used as a sample water body acquisition image sequence for one-time cyclic learning in a model cyclic updating process, so that a plurality of updated training value network units are cyclically updated, and the optimized training image evaluation network is cyclically optimized.

In some exemplary design ideas, after training image evaluation is performed on the second observation sample water body acquisition image through the optimized training image evaluation network to generate an observation description of the second observation sample water body acquisition image, the target water body image can be updated according to the observation description of the second observation sample water body acquisition image and the labeling training value carried by the second observation sample water body acquisition image to generate an adjusted target water body image; updating the sample water body acquisition image sequence according to the adjusted target water body image, and generating an adjusted sample water body acquisition image sequence; the adjusted sample water body acquisition image sequence is used as the sample water body acquisition image sequence of one-time cyclic learning in the model cyclic updating process, so that a plurality of updated training value network units are cyclically updated, and the optimized training image evaluation network is circularly optimized, so that the cyclic optimization of the optimized training image evaluation network can be realized through the step-wise updating of the sample water body acquisition image sequence, and further, the reliability of the training image evaluation network after cyclic optimization on the images with high training values is higher.

In an exemplary design concept, for step S200, it may be implemented by the following steps.

Step W102, acquiring a template water body acquisition image sequence based on the sample water body acquisition image sequence; the template water body acquisition image sequence comprises a plurality of template water body acquisition images corresponding to the water body pollution characteristic detection dimensions.

The template water body acquisition image sequence comprises a plurality of template water body acquisition images corresponding to the water body pollution characteristic detection dimensions. The template water body acquisition images corresponding to the water body pollution characteristic detection dimension are multiple. The dimension of water pollution characteristic detection is used for detecting the water pollution characteristic of the water body acquired image. The water pollution characteristic detection dimension comprises at least one water pollution characteristic detection batch, one water pollution characteristic detection batch corresponds to one water pollution class, and different water pollution characteristic detection batches correspond to different water pollution classes.

Wherein at least one different water pollution characteristic detection batch is included between different water pollution characteristic detection dimensions, for example, water pollution characteristic detection dimension 1 includes water pollution characteristic detection batch a and water pollution characteristic detection batch b, and water pollution characteristic detection dimension 2 includes water pollution characteristic detection batch a and water pollution characteristic detection batch c. The water pollution characteristic detection dimension 1 comprises a water pollution characteristic detection batch a and a water pollution characteristic detection batch b, and the water pollution characteristic detection dimension 2 comprises a water pollution characteristic detection batch c and a water pollution characteristic detection batch d.

Step W104, acquiring a current template water body acquisition image corresponding to the current water body pollution characteristic detection dimension, and updating weight parameters of an image semantic extraction network of the first water body pollution characteristic learning model according to the current template water body acquisition image to generate a second water body pollution characteristic learning model.

The current water pollution characteristic detection dimension can be any water pollution characteristic detection dimension in the template water body acquisition image sequence. The current template water body acquisition image refers to a template water body acquisition image corresponding to the current water body pollution characteristic detection dimension. The first water pollution characteristic learning model refers to a water pollution characteristic learning model of which the model weight information is initialized. The water pollution characteristic learning model is used for detecting the water pollution characteristic of the loaded image and determining a neural network of the water pollution type corresponding to the loaded image. The second water pollution characteristic learning model is a neural network obtained by updating weight parameters of the first water pollution characteristic learning model according to the current template water acquisition image and updating the weight parameters of the image semantic extraction network.

The water pollution characteristic learning model comprises an image semantic extraction network and a characteristic prediction network. The image semantic extraction network is used for extracting image semantics of the loaded image, the feature prediction network is used for predicting the water pollution of the image semantic features of the image semantic extraction network, and finally, the water pollution prediction result corresponding to the loaded image can be generated according to the prediction result of the feature prediction network. The method comprises the steps that a weight parameter is updated on an image semantic extraction network of a water pollution characteristic learning model according to a template water body acquisition image corresponding to a water pollution characteristic detection dimension, so that the water pollution characteristic learning model can learn relevant pollution characteristic knowledge of the water pollution characteristic detection dimension, the trained water pollution characteristic learning model can be used for processing the water pollution characteristic detection dimension, and a target water pollution category corresponding to a loaded image is determined from various water pollution categories corresponding to the water pollution characteristic detection dimension.

In an exemplary design concept, a current water pollution feature detection dimension may be determined from the template water body acquisition image sequence, a current template water body acquisition image corresponding to the current water pollution feature detection dimension is obtained, a weight parameter update is performed on an image semantic extraction network of a first water pollution feature learning model according to the current template water body acquisition image, and weight information of the image semantic extraction network of the first water pollution feature learning model is adjusted, so that a second water pollution feature learning model is obtained.

And step W106, acquiring a migration template water body acquisition image corresponding to the next water body pollution characteristic detection dimension, and updating weight parameters of an image semantic extraction network of the second water body pollution characteristic learning model according to the migration template water body acquisition image to generate a third water body pollution characteristic learning model.

The next water pollution characteristic detection dimension may be other water pollution characteristic detection dimensions except the current water pollution characteristic detection dimension, in other words, the water pollution characteristic detection dimension which does not participate in the water pollution characteristic learning in the template water acquisition image sequence. The migration template water body acquisition image refers to a template water body acquisition image corresponding to the detection dimension of the next water body pollution characteristic. The third water pollution characteristic learning model is a neural network obtained by updating weight parameters of the second water pollution characteristic learning model according to the migration template water body acquisition image and updating the weight parameters of the image semantic extraction network. It can be understood that when the weight parameter updating is performed on the image semantic extraction network, only the weight information of the image semantic extraction network is updated, and when the weight parameter optimization is performed on the feature prediction network subsequently, only the weight information of the feature prediction network is updated.

In an exemplary design concept, a next water pollution feature detection dimension may be determined from the template water body acquisition image sequence, a migration template water body acquisition image corresponding to the next water pollution feature detection dimension is acquired, a weight parameter update is performed on an image semantic extraction network of the second water pollution feature learning model according to the migration template water body acquisition image, and a third water pollution feature learning model is generated, in other words, on the basis of performing weight parameter update and training completion on the image semantic extraction network of the first water pollution feature learning model according to the current template water body acquisition image, the weight parameter update is continuously performed on the image semantic extraction network of the first water pollution feature learning model according to the migration template water body acquisition image corresponding to the next water pollution feature detection dimension.

In an exemplary design concept, updating the weight parameters of the image semantic extraction network of the water pollution feature learning model according to the template water body acquisition image can specifically be that the template water body acquisition image is input into the water pollution feature learning model, the water pollution feature learning model outputs water pollution estimation features corresponding to the template water body acquisition image, a water pollution prediction error value is calculated according to the water pollution labeling features and the water pollution estimation features carried by the template water body acquisition image, and the weight parameters of the image semantic extraction network of the water pollution feature learning model are updated according to the water pollution prediction error value until the water pollution prediction error value converges, so as to generate the water pollution feature learning model of the trained image semantic extraction network.

And step W108, loading the migration training images of the migration template water body acquisition images to the image semantic extraction network of the second water body pollution characteristic learning model and the third water body pollution characteristic learning model respectively, and generating corresponding first water body pollution image vectors and second water body pollution image vectors.

The migration training image of the migration template water body acquisition image refers to each template water body acquisition image which is used for model learning before the migration template water body acquisition image.

The first water pollution image vector refers to data obtained by an image semantic extraction network of a second water pollution characteristic learning model after a migration training image of a migration template water body acquisition image is input into the second water pollution characteristic learning model. The second water pollution image vector refers to characteristic data generated by an image semantic extraction network of a third water pollution characteristic learning model after a migration training image of a migration template water body acquisition image is input into the third water pollution characteristic learning model.

In an exemplary design concept, a migration training image of a migration template water body acquisition image may be obtained from the template water body acquisition image sequence, the migration training image of the migration template water body acquisition image is respectively loaded to image semantic extraction networks of a second water body pollution feature learning model and a third water body pollution feature learning model, extraction features generated by the image semantic extraction networks of the second water body pollution feature learning model are used as a first water body pollution image vector, and extraction features generated by the image semantic extraction networks of the third water body pollution feature learning model are used as a second water body pollution image vector. When a plurality of migration template water body acquisition images exist in the migration template water body acquisition images, loading each migration template water body acquisition image to an image semantic extraction network of a second water body pollution characteristic learning model and a third water body pollution characteristic learning model respectively to generate a first water body pollution image vector and a second water body pollution image vector which are respectively corresponding to each migration template water body acquisition image.

And step W110, carrying out weight parameter optimization on the characteristic prediction network of the third water pollution characteristic learning model according to the first water pollution image vector and the second water pollution image vector corresponding to the water acquisition image of the same template, and generating an optimized second water pollution characteristic learning model.

In an exemplary design concept, the feature prediction network of the third water pollution feature learning model can be subjected to weight parameter optimization according to the first water pollution image vector and the second water pollution image vector corresponding to the water acquisition image of the same template, and weight information of the feature prediction network of the third water pollution feature learning model is adjusted, so that a new water pollution feature learning model is obtained. It can be appreciated that the extracted features generated by the migration training image of the migration template water body acquisition image through the second water body pollution feature learning model are more accurate, because the second water body pollution feature learning model is obtained by training according to the migration training image of the migration template water body acquisition image. The extraction features of the migration training image of the migration template water body acquisition image generated through the third water body pollution feature learning model are inaccurate, because the third water body pollution feature learning model is obtained by further updating weight information according to the migration template water body acquisition image on the basis of the second water body pollution feature learning model. At this time, the third water pollution characteristic learning model trains and learns the knowledge of the water body acquisition image of the transfer template through the image semantic extraction network, so that the water body pollution characteristic detection dimension corresponding to the water body acquisition image of the transfer template has better performance, the characteristic prediction network of the third water body pollution characteristic learning model is optimized according to the first water body pollution image vector and the second water body pollution image vector corresponding to the water body acquisition image of the same template, the purpose is to enable the third water body pollution characteristic learning model to maintain the performance of the water body pollution characteristic detection dimension corresponding to the transfer training image, and the new water body pollution characteristic learning model obtained through the characteristic prediction network training can be suitable for the water body pollution characteristic detection dimension corresponding to all the current template water body acquisition images. And replacing the new water pollution characteristic learning model with the previous second water pollution characteristic learning model, in other words, taking the new water pollution characteristic learning model as the optimized second water pollution characteristic learning model.

In an exemplary design concept, the optimization of the weight parameter of the feature prediction network of the third water pollution feature learning model according to the first water pollution image vector and the second water pollution image vector corresponding to the template water body acquisition image may specifically be that the second water pollution image vector corresponding to the template water body acquisition image is used as the input of the image semantic extraction network of the third water pollution feature learning model, the first water pollution image vector corresponding to the template water body acquisition image is used as the feature extraction result of the image semantic extraction network of the third water pollution feature learning model, and the feature extraction of the third water pollution feature learning model is supervised and trained to generate the water pollution feature learning model after the feature prediction network training. The first water pollution image vector corresponding to the template water body acquisition image is the extraction feature of the image semantic extraction network of the second water pollution feature learning model, and the second water pollution image vector corresponding to the template water body acquisition image is the extraction feature of the image semantic extraction network of the third water pollution feature learning model.

In an exemplary design concept, because the water pollution feature learning model has better model weight information through image semantic extraction network training, when the feature prediction network of the water pollution feature learning model is subjected to weight parameter optimization, a part of migration training images can be obtained from all migration training images of the migration template water body acquisition image, the feature prediction network of the third water pollution feature learning model can be subjected to weight parameter optimization according to a small amount of migration training images, the weight parameters of the feature prediction network are updated, and a new water pollution feature learning model is generated.

And step W112, returning to the step of acquiring a migration template water body acquisition image corresponding to the next water body pollution characteristic detection dimension until the model is converged, and obtaining the water body pollution detection model according to a second water body pollution characteristic learning model corresponding to the convergence of the model.

The water pollution detection model is a water pollution characteristic learning model which is finally trained.

The method comprises the steps of carrying out weight parameter optimization on a feature prediction network of a third water pollution feature learning model according to a first water pollution image vector and a second water pollution image vector corresponding to the same template water body acquisition image, after a new water pollution feature learning model is generated, replacing a previous second water pollution feature learning model with the new water pollution feature learning model, taking the new water pollution feature learning model as an optimized second water pollution feature learning model, and returning to the step of acquiring a migration template water body acquisition image corresponding to the next water pollution feature detection dimension, so as to carry out weight parameter updating of a new round of image semantic extraction network and weight parameter updating of the feature prediction network. The method comprises the steps of updating weight parameters of an image semantic extraction network of a current latest second water pollution feature learning model according to template water collection images corresponding to new next water pollution feature detection dimensions, generating a new third water pollution feature learning model, loading template water collection images corresponding to transfer water pollution feature detection dimensions of the latest trained water pollution feature detection dimensions to the image semantic extraction networks of the current latest second water pollution feature learning model and the third water pollution feature learning model respectively, generating corresponding first water pollution image vectors and second water pollution image vectors, carrying out weight parameter optimization on a feature prediction network of the current latest third water pollution feature learning model according to the first water pollution image vectors and the second water pollution image vectors corresponding to the same template water collection image, generating a new water pollution feature learning model, replacing the previous second water pollution feature learning model with the new water pollution feature learning model, taking the new water pollution feature learning model as the optimized second water pollution feature learning model, returning to the step of acquiring the transfer water pollution feature learning model corresponding to the next water pollution feature detection dimensions, and updating weight parameters of the image data prediction network according to the first water pollution image vectors and the second water pollution image vectors corresponding to the same template water pollution feature detection dimensions, and updating weight parameters of the image prediction network. And similarly, after the template water body acquisition images corresponding to all the water body pollution characteristic detection dimensions in the template water body acquisition image sequence participate in model learning, indicating that training is finished, and obtaining the water body pollution detection model according to the finally obtained water body pollution characteristic learning model when the model converges. Or after the iteration times of the feature prediction network are greater than the set times, the training is stopped, and the water pollution detection model is obtained according to the finally obtained water pollution feature learning model when the model converges. The water pollution detection model can be obtained according to a second water pollution characteristic learning model corresponding to the model convergence, and the finally obtained water pollution characteristic learning model can be used as the water pollution detection model.

The template water body acquisition image sequence comprises a template water body acquisition image 1 corresponding to a water body pollution characteristic detection dimension 1, a template water body acquisition image 2 corresponding to a water body pollution characteristic detection dimension 2, a template water body acquisition image 3 corresponding to a water body pollution characteristic detection dimension 3 and a template water body acquisition image 4 corresponding to a water body pollution characteristic detection dimension 4. And updating weight parameters of an image semantic extraction network of the first water pollution characteristic learning model according to the template water body acquisition image 1 to generate the first water pollution characteristic learning model. And updating weight parameters of an image semantic extraction network of the first water pollution characteristic learning model according to the template water body acquisition image 2 to generate a second water pollution characteristic learning model. Inputting the template water body acquisition image 1 into an image semantic extraction network of a first water body pollution characteristic learning model, generating a first water body pollution image vector of the template water body acquisition image 1 obtained by the image semantic extraction network, inputting the template water body acquisition image 2 into an image semantic extraction network of a second water body pollution characteristic learning model, and generating a second water body pollution image vector of the template water body acquisition image 1 obtained by the image semantic extraction network. And carrying out weight parameter optimization on the characteristic prediction network of the second water pollution characteristic learning model according to the first water pollution image vector and the second water pollution image vector of the template water body acquisition image 1, and generating an optimized second water pollution characteristic learning model.

And updating weight parameters of the image semantic extraction network of the optimized second water pollution characteristic learning model according to the template water body acquisition image 3 to generate a third water pollution characteristic learning model. The template water body acquisition image 1 is respectively loaded to the image semantic extraction network of the optimized second water body pollution characteristic learning model and the optimized third water body pollution characteristic learning model to generate a first water body pollution image vector and a second water body pollution image vector of the template water body acquisition image 1, and the template water body acquisition image 2 is respectively loaded to the image semantic extraction network of the optimized second water body pollution characteristic learning model and the optimized third water body pollution characteristic learning model to generate a first water body pollution image vector and a second water body pollution image vector of the template water body acquisition image 2. And carrying out weight parameter optimization on the characteristic prediction network of the third water pollution characteristic learning model according to the first water pollution image vector and the second water pollution image vector of the template water body acquisition image 1 and the first water pollution image vector and the second water pollution image vector of the template water body acquisition image 2, and generating an optimized third water pollution characteristic learning model.

And updating weight parameters of the image semantic extraction network of the optimized third water pollution characteristic learning model according to the template water body acquisition image 4 to generate a fourth water pollution characteristic learning model. The method comprises the steps of loading template water body acquisition images 1 to image semantic extraction networks of an optimized third water body pollution characteristic learning model and a fourth water body pollution characteristic learning model respectively, generating a first water body pollution image vector and a second water body pollution image vector of the template water body acquisition images 1, loading template water body acquisition images 2 to image semantic extraction networks of the optimized third water body pollution characteristic learning model and the optimized fourth water body pollution characteristic learning model respectively, generating a first water body pollution image vector and a second water body pollution image vector of the template water body acquisition images 2, loading template water body acquisition images 3 to image semantic extraction networks of the optimized third water body pollution characteristic learning model and the optimized fourth water body pollution characteristic learning model respectively, and generating a first water body pollution image vector and a second water body pollution image vector of the template water body acquisition images 3. And carrying out weight parameter optimization on a feature prediction network of a fourth water pollution feature learning model according to the first water pollution image vector and the second water pollution image vector of the template water body acquisition image 1, the first water pollution image vector and the second water pollution image vector of the template water body acquisition image 2 and the first water pollution image vector and the second water pollution image vector of the template water body acquisition image 3 to generate an optimized fourth water pollution feature learning model, and taking the optimized fourth water pollution feature learning model as the water pollution detection model.

Therefore, through the alternate training of the image semantic extraction network and the feature prediction network, the water pollution feature learning model can keep the performance in the water pollution feature detection dimension on the basis of learning the knowledge of the new water pollution feature detection dimension each time. Then, after being loaded into the water pollution detection model, the water pollution detection model can determine the target water pollution type corresponding to the target water acquisition image from the water pollution types corresponding to all trained water pollution characteristic detection dimensions.

Based on the steps, the template water body acquisition image sequence is obtained, the template water body acquisition image sequence comprises a plurality of template water body acquisition images corresponding to water body pollution feature detection dimensions, the current template water body acquisition image corresponding to the current water body pollution feature detection dimensions is obtained, weight parameter updating is carried out on an image semantic extraction network of a first water body pollution feature learning model according to the current template water body acquisition image, a second water body pollution feature learning model is generated, a migration template water body acquisition image corresponding to the next water body pollution feature detection dimensions is obtained, weight parameter updating is carried out on the image semantic extraction network of the second water body pollution feature learning model according to the migration template water body acquisition image, a third water body pollution feature learning model is generated, migration training images of the migration template water body acquisition image are respectively loaded to the image semantic extraction networks of the second water body pollution feature learning model and the third water body pollution feature learning model, corresponding first water body pollution image vectors and second water body pollution image vectors are generated, weight parameter optimization is carried out on a feature prediction network of the third water body pollution feature learning model according to the first water body pollution image corresponding to the same template water body acquisition image, and the migration training images corresponding to the second water body pollution feature learning model are obtained after the migration training images are converged under the condition that the condition of the water body pollution feature learning model is detected, and the water body pollution feature is detected according to the water body pollution feature of the water body model is obtained. Therefore, the training of the image semantic extraction network for the water pollution characteristic learning model can enable the water pollution characteristic learning model to learn the characteristics of each water pollution characteristic detection dimension from the template water body acquisition image corresponding to each water pollution characteristic detection dimension, the training of the characteristic prediction network for the water pollution characteristic learning model can enable the water pollution characteristic learning model to learn the characteristics common to each water pollution characteristic detection dimension, and through the alternate training of the image semantic extraction network and the characteristic prediction network, the water pollution characteristic extraction precision of the water pollution characteristic learning model on the migration water pollution characteristic detection dimension can be ensured, and the water pollution characteristic extraction precision of the water pollution characteristic learning model on the new water pollution characteristic detection dimension can be ensured, so that the common water pollution detection model is obtained.

In an exemplary design concept, acquiring the template water body acquisition image sequence includes:

step W402, acquiring a plurality of sample water body acquisition images; the sample water body acquisition image carries a labeling pollution source.

And step W404, clustering the collected images of each sample water body corresponding to the same labeling pollution source to generate a basic image cluster corresponding to each labeling pollution source.

The labeling pollution source is used for identifying the water pollution type of the sample water body acquisition image.

In an exemplary design concept, a plurality of sample water body acquisition images carrying labeling pollution sources can be obtained, each sample water body acquisition image corresponding to the same labeling pollution source is clustered, and a basic image cluster corresponding to each labeling pollution source is generated. In other words, the computer device may perform the water pollution feature detection on the sample water body collected image based on the water pollution type of the sample water body collected image, and form each sample water body collected image of the same water pollution type into a basic image cluster, so as to generate the basic image clusters corresponding to each water body pollution type respectively.

And step W406, determining training sequence numbers corresponding to the acquired images of the water bodies of the samples according to the labeling pollution sources.

Step W408, the basic image clusters are used as water pollution characteristic detection batches, each water pollution characteristic detection batch corresponding to the same training sequence number is clustered, and a target water pollution characteristic detection image set corresponding to each training sequence number is generated; the target water pollution characteristic detection image set comprises all template water acquisition images corresponding to the same water pollution characteristic detection dimension.

And step W410, obtaining the template water body acquisition image sequence according to each target water body pollution characteristic detection image set.

In an exemplary design concept, the water pollution detection model is used for determining target water pollution detection characteristics corresponding to the target water collection image from the labeling pollution sources corresponding to the template water collection image.

In this embodiment, the first clustering is performed on the sample water body collected images according to the labeling pollution sources, so as to obtain a water body pollution feature detection batch, the second clustering is performed on the water body pollution feature detection batch according to the training sequence number, so as to obtain water body pollution feature detection dimensions, and the template water body collected image sequence for training the general water body pollution feature learning model can be obtained according to the template water body collected images corresponding to each water body pollution feature detection dimension.

In an exemplary design concept, the template water body collection image in the template water body collection image sequence carries a water body pollution labeling feature, the neural network model to be optimized is a first water body pollution feature learning model or a second water body pollution feature learning model, and the updating of the weight parameters of the image semantic extraction network of the neural network model to be optimized comprises the following steps: carrying out image analysis on the loaded template water body acquisition image through the neural network model to be optimized, and generating water body pollution estimation characteristics corresponding to the loaded template water body acquisition image; and updating the weight parameters of the image semantic extraction network of the neural network model to be optimized according to the water pollution labeling features and the loss function values of the water pollution estimation features corresponding to the loaded template water body acquisition image until the loss function values meet the preset requirements.

The water pollution labeling features represent actual water pollution types of the template water body acquired images. The water pollution estimation features represent the estimated water pollution types of the template water acquisition images, namely, the water pollution features output by the water pollution feature learning model are obtained after the template water acquisition images are input into the water pollution feature learning model. The weight parameters of the image semantic extraction network refer to the weight parameters of the image semantic extraction network of the water pollution characteristic learning model.

The method for updating the weight parameters of the image semantic extraction network of the neural network model to be optimized specifically can be that the loaded template water body acquisition image is subjected to image analysis through the neural network model to be optimized to generate water body pollution estimation characteristics corresponding to the loaded template water body acquisition image, a loss function value is calculated according to the water body pollution labeling characteristics and the water body pollution estimation characteristics corresponding to the loaded template water body acquisition image, reverse transfer is carried out according to the loss function value, the weight parameters of the image semantic extraction network of the neural network model to be optimized are updated, in other words, the water body pollution prediction error value is calculated according to the water body pollution labeling characteristics and the water body pollution estimation characteristics corresponding to the loaded template water body acquisition image, and the weight parameters of the image semantic extraction network of the neural network model to be optimized are updated according to the water body pollution prediction error value until the loss function value meets preset requirements, and the water body pollution characteristic learning model of the trained image semantic extraction network is generated.

When the neural network model to be optimized is a first water pollution characteristic learning model, inputting a current template water body acquisition image into the first water pollution characteristic learning model to generate a water pollution estimation characteristic corresponding to the current template water body acquisition image, and adjusting the weight parameters of an image semantic extraction network of the first water pollution characteristic learning model according to the water pollution labeling characteristic and the loss function value of the water pollution estimation characteristic corresponding to the current template water body acquisition image until the loss function value meets the preset requirement, so as to generate a second water pollution characteristic learning model.

When the neural network model to be optimized is a second water pollution characteristic learning model, inputting the migration template water body acquisition image into the second water pollution characteristic learning model to generate a water pollution estimation characteristic corresponding to the migration template water body acquisition image, and adjusting the weight parameter of an image semantic extraction network of the second water pollution characteristic learning model according to the water pollution labeling characteristic corresponding to the migration template water body acquisition image and the loss function value of the water pollution estimation characteristic until the loss function value meets the preset requirement, so as to generate a third water pollution characteristic learning model.

Carrying out image analysis on the loaded template water body acquisition image through the neural network model to be optimized, generating water body pollution estimation characteristics corresponding to the loaded template water body acquisition image, updating the weight parameters of the image semantic extraction network of the neural network model to be optimized according to the water body pollution labeling characteristics corresponding to the loaded template water body acquisition image and the loss function values of the water body pollution estimation characteristics until the loss function values are converged, and generating a water body pollution characteristic learning model of the image semantic extraction network of the trained image. The weight information of the image semantic extraction network can be updated rapidly according to the supervised training, and the weight parameters of the image semantic extraction network of the water pollution characteristic learning model are updated rapidly.

In an exemplary design concept, the current water pollution feature learning model is any one of a first water pollution feature learning model, a second water pollution feature learning model, a third water pollution feature learning model and the water pollution detection model, an image semantic extraction network of the current water pollution feature learning model is used for carrying out image semantic extraction on a loaded template water acquisition image to generate image semantic features, and a feature prediction network of the current water pollution feature learning model is used for carrying out water pollution pixel feature prediction on the image semantic features to generate water pollution pixel features.

The image semantic extraction network of the current water pollution feature learning model is used for carrying out image semantic extraction on the loaded template water body acquisition image to generate image semantic features, the feature prediction network is used for carrying out water pollution pixel feature prediction on the image semantic features to generate water pollution pixel features, and finally, the water pollution estimated features corresponding to the loaded template water body acquisition image can be obtained according to the water pollution pixel features obtained by the feature prediction network. The current water pollution characteristic learning model is any one of a first water pollution characteristic learning model, a second water pollution characteristic learning model, a third water pollution characteristic learning model and the water pollution detection model. It will be appreciated that the current water pollution feature learning model refers to the currently used water pollution feature learning model. In other words, all the water pollution feature learning models of the application carry out image semantic extraction on the loaded template water body acquisition image through the image semantic extraction network to generate image semantic features, and carry out water pollution pixel feature prediction on the image semantic features through the feature prediction network to generate water pollution pixel features.

In an exemplary design concept, an image semantic extraction network of a water pollution feature learning model includes at least one convolution unit, and a feature prediction network includes at least one feature prediction branch, where one convolution unit corresponds to one feature prediction branch. When the image semantic extraction network of the water pollution characteristic learning model comprises a plurality of convolution units, each convolution unit is used for extracting different convolution characteristics of the loaded template water body acquisition image.

The image semantic extraction network of the water pollution feature learning model is used for carrying out image semantic extraction on the loaded template water body acquisition image to generate image semantic features, and the feature prediction network of the water pollution feature learning model is used for carrying out water pollution pixel feature prediction on the image semantic features transmitted by the image semantic extraction network to generate water pollution pixel features. Through cooperation of the image semantic extraction network and the feature prediction network, the water pollution feature learning model can output water pollution features corresponding to the loaded image.

In an exemplary design concept, a loaded template water body acquisition image comprises foreground image data and background image data, an image semantic extraction network comprises a convolution unit corresponding to the foreground image data and a convolution unit corresponding to the background image data, and the extraction of image semantic features comprises the following steps:

Step W502, performing image semantic extraction on the foreground image data according to the convolution unit corresponding to the foreground image data, and generating foreground image semantic features.

And step W504, performing image semantic extraction on the background image data according to the convolution unit corresponding to the background image data to generate background image semantic features.

And step W506, obtaining image semantic features according to the background image semantic features and the foreground image semantic features.

The loaded template water body acquisition image comprises foreground image data and background image data, and the image semantic extraction network comprises a convolution unit corresponding to the foreground image data and a convolution unit corresponding to the background image data. After the loaded template water body acquisition image is input into the water body pollution feature learning model, the foreground image data in the loaded template water body acquisition image can be subjected to image semantic extraction through a convolution unit corresponding to the foreground image data to generate foreground image semantic features corresponding to the foreground image data, the background image data in the loaded template water body acquisition image is subjected to image semantic extraction through a convolution unit corresponding to the background image data to generate background image semantic features corresponding to the background image data, and finally, the background image semantic features and the foreground image semantic features form image semantic features.

In an exemplary design concept, a current water pollution feature learning model comprises a background extraction network, an image semantic extraction network comprises a convolution unit corresponding to the background extraction network and a convolution unit corresponding to a loaded template water body acquisition image, and the extraction of the image semantic features comprises the following steps: carrying out background extraction on the loaded template water body acquisition image through a background extraction network to generate background image data corresponding to the loaded template water body acquisition image; performing image semantic extraction on the background image data through convolution units corresponding to the background extraction network to generate background image semantic features; carrying out image semantic extraction on the loaded template water body acquisition image through a convolution unit corresponding to the loaded template water body acquisition image to generate a foreground image semantic feature; and obtaining image semantic features according to the background image semantic features and the foreground image semantic features.

Illustratively, the water pollution feature learning model may include a background extraction network in addition to the image semantic extraction network and the feature prediction network. After the template water body acquisition image is input into the water body pollution characteristic learning model, the background extraction network can firstly carry out background extraction on the loaded template water body acquisition image to generate background image data corresponding to the loaded template water body acquisition image. Then, respectively loading the background image data and the loaded template water body acquisition image to the corresponding convolution units, carrying out image semantic extraction on the background image data through the convolution units corresponding to the background extraction network to generate background image semantic features, carrying out image semantic extraction on the loaded template water body acquisition image through the convolution units corresponding to the loaded template water body acquisition image to generate foreground image semantic features, and finally forming image semantic features by the background image semantic features and the foreground image semantic features.

In an exemplary design concept, the background image data includes at least one background image block, the background image semantic feature includes a background image block vector corresponding to each background image block, and the image semantic feature is obtained according to the background image semantic feature and the foreground image semantic feature, including:

and step W702, performing salient feature distribution on the background image block vectors corresponding to the background image blocks according to the semantic features of the foreground images, and generating salient feature information corresponding to the background image blocks.

The saliency feature allocation is used for allocating different saliency feature information to each background image block, and the saliency feature information of the background image blocks can reflect the importance of the background image blocks in the image.

Step W704, regularized conversion is carried out on the salient feature information corresponding to each background image block, and weight information corresponding to each background image block is generated.

In an exemplary design concept, saliency feature information corresponding to each background image block may be regularly converted to generate weight information corresponding to each background image block. In other words, the saliency feature information corresponding to each background image block may be subjected to regularization conversion to generate weights corresponding to each background image block.

And step W706, obtaining updated background image semantic features according to the background image block vectors and the weight information corresponding to the background image blocks.

Step W708, obtaining image semantic features according to the foreground image semantic features and the updated background image semantic features.

For example, when determining updated background image semantic features, the computer device may combine the foreground image semantic features and the updated background image semantic features into image semantic features. Further, the image semantic features are loaded to a feature prediction network, water pollution pixel feature prediction is carried out on the foreground image semantic features through feature prediction branches corresponding to the foreground image data, a water pollution image extraction vector corresponding to the foreground image data is generated, water pollution pixel feature prediction is carried out on the updated background image semantic features through feature prediction branches corresponding to the background image data, a water pollution image extraction vector corresponding to the background image data is generated, and the water pollution pixel features are formed by the water pollution image extraction vector corresponding to the foreground image data and the water pollution image extraction vector corresponding to the background image data.

In an exemplary design concept, an image semantic extraction network comprises a plurality of convolution units, a feature prediction network comprises a plurality of feature prediction branches, the convolution units are in one-to-one correspondence with the feature prediction branches, the image semantic features comprise initial image extraction vectors generated by the convolution units, and the extraction of the water pollution pixel features comprises the following steps: loading the initial image extraction vectors generated by each convolution unit to the corresponding feature prediction branches to generate water pollution image extraction vectors output by each feature prediction branch; and extracting vectors according to each water pollution image to obtain the water pollution pixel characteristics.

For example, the image semantic extraction network of the water pollution feature learning model may include a plurality of convolution units, and the feature prediction network may include a plurality of feature prediction branches, one convolution unit corresponding to one feature prediction branch, in other words, the convolution unit and the feature prediction branch are in one-to-one correspondence. After the template water body acquisition image is loaded to the water body pollution characteristic learning model, each convolution unit can generate a corresponding initial image extraction vector through image analysis of each convolution unit. Because one convolution unit corresponds to one feature prediction branch, the initial image extraction vector generated by each convolution unit is loaded into the corresponding feature prediction branch, and each feature prediction branch can generate the corresponding water pollution image extraction vector through the image analysis of each feature prediction branch, and the water pollution pixel feature is obtained through each water pollution image extraction vector.

In an exemplary design concept, a feature prediction network of a third water pollution feature learning model is optimized according to a first water pollution image vector and a second water pollution image vector corresponding to a water acquisition image of the same template, and an optimized second water pollution feature learning model is generated, which comprises: inputting a second water pollution image vector corresponding to the migration training image into a characteristic prediction network of a third water pollution characteristic learning model to generate a corresponding water pollution estimation characteristic; calculating a target water pollution prediction error value according to a first water pollution image vector and water pollution estimation characteristics corresponding to the water acquisition image of the same template; and updating the weight information of the characteristic prediction network of the third water pollution characteristic learning model according to the target water pollution prediction error value until the loss function value meets the preset requirement, and generating an optimized second water pollution characteristic learning model.

For example, in order to enable the third water pollution feature learning model to maintain the performance on the migration training image of the migration template water acquisition image after training by the migration template water acquisition image, the model is required to perform weight parameter optimization on the feature prediction network of the third water pollution feature learning model. The second water pollution image vector can be used as the input of the characteristic prediction network of the third water pollution characteristic learning model, the first water pollution image vector is used as the expected output of the characteristic prediction network of the third water pollution characteristic learning model, and the characteristic prediction network of the third water pollution characteristic learning model is subjected to weight parameter optimization, so that the third water pollution characteristic learning model can keep the performance on the migration training image of the migration template water body acquisition image.

The second water pollution image vector corresponding to the migration training image can be input into the feature prediction network of the third water pollution feature learning model to generate a water pollution estimation feature obtained by the feature prediction network, a target water pollution prediction error value is calculated according to the first water pollution image vector corresponding to the water acquisition image of the same template and the water pollution estimation feature, reverse transmission is carried out according to the target water pollution prediction error value, weight information of the feature prediction network of the third water pollution feature learning model is updated, training is continued until the loss function value meets the preset requirement, and the water pollution feature learning model of the trained feature prediction network is generated. And taking the water pollution characteristic learning model of the trained characteristic prediction network as a new second water pollution characteristic learning model to replace the previous second water pollution characteristic learning model.

In an exemplary design concept, an image semantic extraction network includes a plurality of convolution units, a feature prediction network includes a plurality of feature prediction branches, the convolution units are in one-to-one correspondence with the feature prediction branches, a first water pollution image vector includes a first image vector generated by each convolution unit of a second water pollution feature learning model, a second water pollution image vector includes a second image vector generated by each convolution unit of a third water pollution feature learning model, and a weight parameter optimization is performed on a feature prediction network of the third water pollution feature learning model according to the first water pollution image vector and the second water pollution image vector corresponding to a water acquisition image of the same template, so as to generate an optimized second water pollution feature learning model, including:

and step W802, inputting each second image vector corresponding to the migration training image into a corresponding feature prediction branch in the third water pollution feature learning model to generate each corresponding water pollution feature.

The image semantic extraction network of the third water pollution feature learning model comprises a plurality of convolution units, and the feature prediction network comprises a plurality of feature prediction branches, wherein one convolution unit corresponds to one feature prediction branch. The template water body acquisition images are respectively loaded to image semantic extraction networks of a second water body pollution characteristic learning model and a third water body pollution characteristic learning model, each convolution unit of the second water body pollution characteristic learning model can output first image vectors corresponding to the template water body acquisition images, each first image vector combination obtains first water body pollution image vectors, each convolution unit of the third water body pollution characteristic learning model can output second image vectors corresponding to the template water body acquisition images, and each second image vector combination obtains second water body pollution image vectors. Therefore, each second image vector corresponding to the migration training image can be input into the corresponding feature prediction branch in the third water pollution feature learning model, and each feature prediction branch outputs the corresponding water pollution feature.

Step W804, a convolution unit and a feature prediction branch having a mapping relationship are acquired as a target convolution unit and a target feature prediction branch.

And step W806, calculating a water pollution prediction error value according to the first image vector generated by the target convolution unit corresponding to the water acquisition image of the same template and the water pollution characteristics output by the target characteristic prediction branch, and generating the water pollution prediction error value corresponding to each characteristic prediction branch.

Illustratively, because one convolution unit corresponds to one feature prediction branch, one first image vector corresponds to one water pollution feature. Then, the water pollution prediction error value can be calculated according to the first image vector generated by the target convolution unit corresponding to the water acquisition image of the same template and the water pollution characteristics output by the target characteristic prediction branch, so as to generate the water pollution prediction error value respectively corresponding to each characteristic prediction branch. The target convolution unit and the target feature prediction branch refer to a convolution unit and a feature prediction branch with a mapping relation.

And step W808, updating the weight parameters of the corresponding feature prediction branches in the third water pollution feature learning model according to the water pollution prediction error values until the loss function value meets the preset requirement, and generating an optimized second water pollution feature learning model.

In an exemplary design concept, the water pollution prediction error values can be reversely transferred, and the weight parameters of the corresponding feature prediction branches in the third water pollution feature learning model are respectively updated and continuously trained until the loss function value meets the preset requirement, so that a water pollution feature learning model of the trained feature prediction network is generated. And taking the water pollution characteristic learning model of the trained characteristic prediction network as a new second water pollution characteristic learning model to replace the previous second water pollution characteristic learning model.

The third water pollution characteristic learning model comprises a convolution unit I with a mapping relation, a characteristic prediction branch I, a convolution unit II with a mapping relation and a characteristic prediction branch II. The template water body acquisition image can be respectively loaded to the image semantic extraction network of the second water body pollution characteristic learning model and the image semantic extraction network of the third water body pollution characteristic learning model, the first convolution unit of the second water body pollution characteristic learning model outputs a first image vector a1, the second convolution unit of the second water body pollution characteristic learning model outputs a first image vector a2, the first convolution unit of the third water body pollution characteristic learning model outputs a second image vector b1, and the second convolution unit of the third water body pollution characteristic learning model outputs a second image vector b2. Then, the second image vector b1 may be loaded to the first feature prediction branch of the third water pollution feature learning model, a first image extraction vector c1 generated by the first feature prediction branch is generated, a water pollution prediction error value d1 is calculated according to the first image vector a1 and the first image extraction vector c1, and weight information of the first feature prediction branch of the third water pollution feature learning model is adjusted according to the water pollution prediction error value d 1. Then, the second image vector b2 is loaded to a second feature prediction branch of the third water pollution feature learning model to generate a first image extraction vector c2, a water pollution prediction error value d2 is calculated according to the first image vector a2 and the first image extraction vector c2, and weight information of the second feature prediction branch of the third water pollution feature learning model is adjusted according to the water pollution prediction error value d 2. After the weight information of the feature prediction branch is updated, a new water pollution prediction error value d1 and a new water pollution prediction error value d2 can be obtained after the operation is circulated, and when the updated water pollution prediction error value d1 and the updated water pollution prediction error value d2 are converged, a water pollution feature learning model of the trained feature prediction network can be generated.

Fig. 2 illustrates a hardware structural intent of a visible light image analysis-based water feature detection system 100 for implementing the visible light image analysis-based water feature detection method according to an embodiment of the present application, as shown in fig. 2, the visible light image analysis-based water feature detection system 100 may include a processor 110, a machine-readable storage medium 120, a bus 130, and a communication unit 140.

In some alternative embodiments, the visible light image analysis-based water feature detection system 100 may be a single server or a group of servers. The server farm may be centralized or distributed (e.g., the visible light image analysis-based water feature detection system 100 may be a distributed system). In some alternative embodiments, the visible light image analysis-based water feature detection system 100 may be local or remote. For example, the visible light image analysis-based water feature detection system 100 may access information and/or data stored in the machine-readable storage medium 120 via a network. For another example, the visible light image analysis based water feature detection system 100 may be directly connected to the machine-readable storage medium 120 to access stored information and/or data. In some alternative embodiments, the visible light image analysis based water feature detection system 100 may be implemented on a cloud platform. For example only, the cloud platform may include a private cloud, a public cloud, a hybrid cloud, a community cloud, a distributed cloud, an internal cloud, a multi-layer cloud, or the like, or any combination thereof.

The machine-readable storage medium 120 may store data and/or instructions. In a specific implementation, at least one processor 110 executes computer-executable instructions stored by the machine-readable storage medium 120, so that the processor 110 may perform the method for detecting a water feature based on visible light image analysis according to the above method embodiment, where the processor 110, the machine-readable storage medium 120, and the communication unit 140 are connected through the bus 130, and the processor 110 may be used to control the transceiving actions of the communication unit 140.

The specific implementation process of the processor 110 may refer to the above embodiments of the method executed by the water feature detection system 100 based on visible light image analysis, and the implementation principle and technical effects are similar, which are not described herein again.

In addition, the embodiment of the application also provides a readable storage medium, wherein computer executable instructions are preset in the readable storage medium, and when a processor executes the computer executable instructions, the water body characteristic detection method based on visible light image analysis is realized.

Similarly, it should be noted that in order to simplify the description of the present disclosure and thereby aid in understanding one or more embodiments of the application, various features are sometimes grouped together in a single embodiment, figure, or description thereof. Similarly, it should be noted that in order to simplify the description of the present disclosure and thereby aid in understanding one or more embodiments of the application, various features are sometimes grouped together in a single embodiment, figure, or description thereof.

Claims

1. The method for detecting the water body characteristics based on the visible light image analysis is characterized by being realized by a water body characteristic detection system based on the visible light image analysis, and comprises the following steps:

visual information prompt is carried out on the management service terminal corresponding to the target water body area based on the water body pollution detection characteristics of the input water body acquisition image;

the training image evaluation network performs training image evaluation on a plurality of sample water body acquisition image sequences based on the training completion to screen out sample water body acquisition image sequences for learning water body pollution characteristics, and the method comprises the following steps:

Acquiring a plurality of example water body acquisition image sequences, any one of the plurality of example water body acquisition image sequences comprising an active example water body acquisition image, a passive example water body acquisition image, and an unsupervised example water body acquisition image, the passive example water body acquisition image comprising: the method comprises the steps that a target water body image with unmatched observation description and labeling training value in a first observation sample water body acquisition image is observed, the observation description is description information generated by training image evaluation on the target water body image according to a training image evaluation network, and the labeling training value is a training value generated by training value calibration on the target water body image;

performing parameter updating on a plurality of training value network units in the training image evaluation network according to the plurality of sample water body acquisition image sequences to generate target threshold training values of each training value network unit in the plurality of training value network units, determining the target threshold training value of each training value network unit in the plurality of training value network units as the threshold training value of each updated training value network unit, and generating a plurality of updated training value network units; wherein, the different sample water body acquisition image sequences correspond to different training value network units;

Optimizing the threshold training values of the training image evaluation network according to the threshold training values of the plurality of updated training value network units, and generating an optimized training image evaluation network;

performing training image evaluation on a second observation sample water body acquisition image according to the optimized training image evaluation network to generate an observation description of the second observation sample water body acquisition image, wherein the observation description of the second observation sample water body acquisition image is obtained according to a threshold training value of the optimized training image evaluation network;

determining a target training evaluation score indicated by the observation description of the second observation sample water body acquisition image, wherein the target training evaluation score is one training evaluation score in a plurality of training evaluation scores, and different training evaluation scores in the plurality of training evaluation scores correspond to different threshold value intervals in a plurality of threshold value intervals;

and determining a second observation sample water body acquisition image which can be used for carrying out water body pollution characteristic learning based on the training evaluation score, and constructing a sample water body acquisition image sequence.

2. The method for detecting water features based on visible light image analysis according to claim 1, wherein the training image evaluation is performed on a second observation sample water collection image according to the optimized training image evaluation network, and after generating the observation description of the second observation sample water collection image, further comprises:

Adjusting the target water body image according to the observation description of the second observation sample water body acquisition image and the annotation training value carried by the second observation sample water body acquisition image to generate an adjusted target water body image, wherein the observation description of the adjusted target water body image is not matched with the annotation training value of the adjusted target water body image;

adjusting the sample water body acquisition image sequence, wherein the negative sample water body acquisition image included in the adjusted sample water body acquisition image sequence is obtained according to the adjusted target water body image and the annotation training value of the adjusted target water body image;

and taking the adjusted sample water body acquisition image sequence as a sample water body acquisition image sequence which is circularly learned for one time in a model circular updating process, so as to circularly update the plurality of updated training value network units and circularly optimize the optimized training image evaluation network.

3. The method for detecting water body characteristics based on visible light image analysis according to claim 1, wherein the performing parameter updating on the plurality of training value network elements in the training image evaluation network according to the plurality of sample water body acquisition image sequences to generate the target gate training value of each training value network element in the plurality of training value network elements comprises:

Taking the non-supervision example water body acquisition image as a negative example water body acquisition image, carrying out parameter updating on a plurality of training value network units in the training image evaluation network according to the plurality of example water body acquisition image sequences, and calculating network performance indexes of each training value network unit under different threshold training values, wherein the network performance indexes are used for indicating the performance of each training value network unit under different threshold training values;

and determining the threshold training value corresponding to the maximum network performance index in the network performance indexes of each training value network unit as the target threshold training value of each training value network unit in the plurality of training value network units.

4. A method for detecting water body characteristics based on visible light image analysis according to claim 1 or 3, wherein said optimizing the threshold training value of the training image evaluation network according to the threshold training values of the plurality of updated training value network elements, generating an optimized training image evaluation network, comprises:

processing the threshold training values of the plurality of updated training value network units to generate the threshold training value of the optimized training image evaluation network;

And generating the optimized training image evaluation network according to the plurality of updated training value network units and the threshold training value of the optimized training image evaluation network.

5. The method for detecting water body characteristics based on visible light image analysis according to claim 4, wherein the processing the threshold training values of the plurality of updated training value network elements to generate the threshold training value of the optimized training image evaluation network comprises:

processing the threshold training values of the plurality of updated training value network units to generate a reference threshold training value;

determining the threshold training value of the optimized training image evaluation network, wherein the number of the threshold training values of the optimized training image evaluation network is multiple, the threshold training value of the optimized training image evaluation network at least comprises the reference threshold training value, the multiple threshold training values of the optimized training image evaluation network form multiple threshold value intervals, and different threshold value intervals correspond to different observation descriptions.

6. The method for detecting water features based on visible light image analysis according to claim 1, wherein the training image evaluation is performed on a second observation sample water acquisition image according to the optimized training image evaluation network, and generating an observation description of the second observation sample water acquisition image includes:

Performing training value observation on the second observation example water body acquisition image according to each updated training value network element in the optimized training image evaluation network, generating a plurality of training value observations of the second observation example water body acquisition image, wherein one training value observation of the plurality of training value observations of the second observation example water body acquisition image corresponds to one updated training value network element, and the plurality of training value observations of the second observation example water body acquisition image comprises one or more of the following: positive results and negative results;

generating a global training value observation of the second observation example water body acquisition image according to the plurality of training value observations of the second observation example water body acquisition image, wherein the global training value observation of the second observation example water body acquisition image is used for reflecting the ratio of positive results in the plurality of training value observations of the second observation example water body acquisition image to the plurality of training value observations of the second observation example water body acquisition image;

and determining an observation description corresponding to a target threshold value interval to which the global training value observation result of the second observation sample water body acquisition image belongs as the observation description of the second observation sample water body acquisition image, wherein the target threshold value interval is one threshold value interval of a plurality of threshold value intervals determined according to the threshold training value of the optimized training image evaluation network.

7. The method of claim 1, wherein the acquiring a plurality of sample water acquisition image sequences comprises:

acquiring a reference water body acquisition image sequence comprising a plurality of supervised active exemplar images, any one of the plurality of supervised active exemplar images comprising source acquisition vectors of a training valid image source, and a plurality of non-supervised exemplar images comprising source acquisition vectors of an image source of unknown training validity;

constructing the plurality of sample water body acquisition image sequences, wherein positive sample water body acquisition images in any one of the sample water body acquisition image sequences are randomly extracted from a plurality of supervised positive sample images in the reference water body acquisition image sequence, non-supervised sample water body acquisition images in any one of the sample water body acquisition image sequences are randomly extracted from a plurality of non-supervised sample images in the reference water body acquisition image sequence, and negative sample water body acquisition images in any one of the sample water body acquisition image sequences are acquired after training image evaluation of the first observation sample water body acquisition image according to the training image evaluation network.

8. The method for detecting water body characteristics based on visible light image analysis according to claim 1, wherein the step of training a water body pollution characteristic learning model based on the sample water body acquisition image sequence to obtain a trained water body pollution detection model comprises the following steps:

acquiring a template water body acquisition image sequence based on the sample water body acquisition image sequence; the template water body acquisition image sequence comprises a plurality of template water body acquisition images corresponding to the water body pollution characteristic detection dimensions;

acquiring a current template water body acquisition image corresponding to a current water body pollution characteristic detection dimension, and updating weight parameters of an image semantic extraction network of a first water body pollution characteristic learning model according to the current template water body acquisition image to generate a second water body pollution characteristic learning model;

acquiring a migration template water body acquisition image corresponding to the next water body pollution characteristic detection dimension, and updating weight parameters of an image semantic extraction network of a second water body pollution characteristic learning model according to the migration template water body acquisition image to generate a third water body pollution characteristic learning model;

respectively loading transfer training images of the transfer template water body acquisition images to image semantic extraction networks of a second water body pollution characteristic learning model and a third water body pollution characteristic learning model to generate corresponding first water body pollution image vectors and second water body pollution image vectors;

Carrying out weight parameter optimization on a characteristic prediction network of a third water pollution characteristic learning model according to a first water pollution image vector and a second water pollution image vector corresponding to the water acquisition image of the same template, and generating an optimized second water pollution characteristic learning model;

and returning to the step of acquiring the migration template water body acquisition image corresponding to the next water body pollution characteristic detection dimension until the model converges, and obtaining the water body pollution detection model according to the second water body pollution characteristic learning model corresponding to the model convergence.

9. A visible light image analysis-based water feature detection system, comprising a processor and a machine-readable storage medium having stored therein machine-executable instructions loaded and executed by the processor to implement the visible light image analysis-based water feature detection method of any one of claims 1-8.