CN116434076A - Remote sensing image target recognition method integrating priori knowledge - Google Patents

Remote sensing image target recognition method integrating priori knowledge Download PDF

Info

Publication number
CN116434076A
CN116434076A CN202310280652.4A CN202310280652A CN116434076A CN 116434076 A CN116434076 A CN 116434076A CN 202310280652 A CN202310280652 A CN 202310280652A CN 116434076 A CN116434076 A CN 116434076A
Authority
CN
China
Prior art keywords
target
slice
layer
feature
prediction
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202310280652.4A
Other languages
Chinese (zh)
Inventor
张筱晗
吕亚飞
路遥
田菁
毕瑷鹏
邢相薇
李今飞
齐江帆
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Beijing Institute of Remote Sensing Information
Original Assignee
Beijing Institute of Remote Sensing Information
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Beijing Institute of Remote Sensing Information filed Critical Beijing Institute of Remote Sensing Information
Priority to CN202310280652.4A priority Critical patent/CN116434076A/en
Publication of CN116434076A publication Critical patent/CN116434076A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/10—Terrestrial scenes
    • G06V20/13—Satellite images
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/764—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using classification, e.g. of video objects
    • G06V10/765—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using classification, e.g. of video objects using rules for classification or partitioning the feature space
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
    • G06V10/774—Generating sets of training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
    • G06V10/778—Active pattern-learning, e.g. online learning of image or video features
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
    • G06V10/80—Fusion, i.e. combining data from various sources at the sensor level, preprocessing level, feature extraction level or classification level
    • G06V10/806—Fusion, i.e. combining data from various sources at the sensor level, preprocessing level, feature extraction level or classification level of extracted features
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/82—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using neural networks

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Evolutionary Computation (AREA)
  • Software Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Computing Systems (AREA)
  • Databases & Information Systems (AREA)
  • Multimedia (AREA)
  • General Health & Medical Sciences (AREA)
  • Medical Informatics (AREA)
  • General Engineering & Computer Science (AREA)
  • Astronomy & Astrophysics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Data Mining & Analysis (AREA)
  • Molecular Biology (AREA)
  • Remote Sensing (AREA)
  • Mathematical Physics (AREA)
  • Image Analysis (AREA)

Abstract

The invention discloses a remote sensing image target identification method integrating priori knowledge, which comprises the following steps: preprocessing the acquired satellite remote sensing image set containing the identification target to obtain a training slice information set; training the target recognition initial model fused with priori knowledge by using the training slice information set to obtain a target recognition model; and processing the satellite remote sensing image to be detected by using the target recognition model to obtain an image target class result. In the invention, in the process of preprocessing the satellite remote sensing image set and training the target recognition model, the target recognition initial model recognition training is guided by introducing the prior knowledge accumulated by manual visual interpretation, so that the human-computer effective fusion is realized. According to the invention, the prior knowledge is fused, so that the target recognition precision of the remote sensing image is improved, and the interpretability and the robustness of the target recognition model are enhanced.

Description

Remote sensing image target recognition method integrating priori knowledge
Technical Field
The invention relates to the technical field of remote sensing data processing, in particular to a remote sensing image target identification method integrating priori knowledge.
Background
With the development of space-based remote sensing technology, satellites have become an important means for implementing earth observation and obtaining surface and ground feature information, and target slice interpretation has important significance in various fields. Target recognition serves as a basic task of image interpretation, aims at establishing connection with high-level semantic information of a target category by extracting image space characteristics, and is a precondition for implementing automatic interpretation.
Currently, in the field of computer vision, intelligent methods based on deep learning have made rapid progress, and in some tasks such as image net natural scene image target recognition, machine recognition accuracy has even exceeded that of humans. However, in practical application, satellite image target recognition is still mainly judged by artificial vision, and the intelligent level is relatively low. Compared with a natural scene image, the target slice has the characteristics of over-top imaging, low resolution, sparse target distribution, complex background and relatively less effective information, and the intelligent model has few extractable characteristics, so that challenges are brought to automatic interpretation of the image; in addition, the intelligent model is similar to a black box, the extracted features are poor in interpretability, and difficulty is brought to targeted tuning and improvement of recognition accuracy of the model. Considering that the interpreter accumulates a great deal of experience in the visual interpretation of the remote sensing image, a plurality of basic features such as target size, length-width ratio and the like play an important role in the identification of the manual interpretation target, how to integrate the priori knowledge into the intelligent recognition model, and the intelligent recognition is implemented by guiding the model by using the experience of the interpreter, so that the method has important value in improving the recognition precision of the intelligent model.
Disclosure of Invention
The technical problem to be solved by the invention is to provide the remote sensing image target recognition method integrating the priori knowledge, which can integrate the priori knowledge into the intelligent recognition model, guide the model to implement intelligent recognition by using the priori knowledge, and improve the recognition precision of the intelligent model.
In order to achieve the above purpose, the embodiment of the invention discloses a remote sensing image target identification method integrating priori knowledge, which comprises the following steps:
s1, acquiring a satellite remote sensing image set containing an identification target; the satellite remote sensing image set comprises M satellite remote sensing images; m is a positive integer not less than 1;
s2, preprocessing the satellite remote sensing image set to obtain a training slice information set;
s3, constructing a target identification initial model fused with priori knowledge; the target identification initial model integrating priori knowledge comprises a first network model, a second network model and an integrated prediction model;
s4, training the target recognition initial model fused with priori knowledge by utilizing the training slice information set to obtain a target recognition model;
s5, processing the satellite remote sensing image to be detected by using the target recognition model to obtain an image target class result; the satellite remote sensing image to be detected characterizes a satellite remote sensing image of which the target category needs to be predicted.
As an optional implementation manner, in an embodiment of the present invention, the preprocessing the satellite remote sensing image set to obtain a training slice information set includes:
s21, carrying out noise reduction, image correction and format conversion processing on the satellite remote sensing image set to obtain a first sample set; the first sample includes a target slice; the target slice comprises an image slice containing a target and image space resolution; the target slice is in a ". Tiff" format;
s22, labeling the first sample set according to a preset priori knowledge model to obtain a second sample set; the second sample comprises a target slice and a labeling file; the annotation file comprises a target category and a target priori knowledge geometric feature; the target priori knowledge geometric features comprise a target length, a target width and a target aspect ratio;
s23, filling all target slices in the second sample set to obtain a training slice information set; the training slice information comprises a training slice and a labeling file.
In an optional implementation manner, in an embodiment of the present invention, the filling processing is performed on all target slices in the second sample set to obtain a training slice information set, including:
S231, any target slice in the second sample set is acquired;
s232, respectively adjusting the width and the height of any target slice to a preset width value and a preset height value to obtain target slices with the same width and height; the preset width value is not smaller than the width of the target slice; the preset height value is not smaller than the height of the target slice;
s233, overlapping and superposing the target slices with the same width and height with a preset full-black image center point to obtain a training slice; the preset full-black image represents an image with all pixel values of 0, and the width and the height of the preset full-black image are respectively the preset width value and the preset height value;
s234, updating any target slice in the second sample set to be a training slice;
s235, judging whether all target slices in the second sample set are updated to training slices, and obtaining a first judgment result;
if the first judgment result is negative, updating the target slice which is not updated into any target slice, and triggering and executing the step S232;
and if the first judgment result is yes, the updated second sample set is a training slice information set.
As an optional implementation manner, in an embodiment of the present invention, the constructing an initial model of target identification with integrated a priori knowledge includes:
S31, constructing a first network model; the first network model is used for processing an input slice to obtain a target depth characteristic f i The method comprises the steps of carrying out a first treatment on the surface of the The target depth feature f i ∈d 1×1×2048 ；
S32, constructing a second network model; the second network model is used for processing the input slice to obtain a target geometric characteristic predicted value; the target geometric characteristic predicted value comprises a target width predicted value, a target length predicted value and a target length-width ratio predicted value;
s33, constructing a fusion prediction model; and the fusion prediction model is used for carrying out fusion processing on the target depth characteristic and the target geometric characteristic predicted value to obtain target category predicted information.
As an alternative implementation manner, in an embodiment of the present invention, the first network model is a convolutional neural network VGG, or, res net, or, seNet, or, shuffleNet, or, googleNet.
As an optional implementation manner, in an embodiment of the present invention, the second network model includes a geometric feature extraction network module and a geometric feature prediction module;
the geometric feature extraction network module comprises:
layer 1, input layer, dimension W×H×N; w, H and N respectively represent the width, height and channel number of an input slice;
Layer 2, the convolution layer, comprising 64 convolution kernels of dimension 3 x 3, outputting features of dimension W x H x 64;
layer 3, the convolution layer, comprising 64 convolution kernels of dimension 3×3, outputting features of dimension w×h×64;
layer 4, pooling layer, using 2×2 max pooling, stride=2, feature dimension becomes (W/2) × (H/2) ×64;
layer 5, the convolution layer, comprising 128 convolution kernels of dimension 3 x 3, outputting features of dimension (W/2) x (H/2) x 128;
layer 6, the convolution layer, comprising 128 convolution kernels of dimension 3 x 3, outputting features of dimension (W/2) x (H/2) x 128;
layer 7, pooling layer, using 2×2 max pooling, stride=2, feature dimension becomes (W/4) × (H/4) ×128;
layer 8, the convolution layer, contains 256 convolution kernels with dimensions of 3×3, output features with dimensions of (W/4) × (H/4) ×256;
layer 9, the convolution layer, contains 256 convolution kernels with dimensions of 3×3, output features with dimensions of (W/4) × (H/4) ×256;
layer 10, pooling layer, using 2×2 max pooling, stride=2, feature dimension becomes (W/8) × (H/8) ×256;
layer 11, the convolution layer, contains 256 convolution kernels with dimensions of 3 x 3, output features with dimensions of (W/8) × (H/8) ×256;
Layer 12, the convolution layer, contains 256 convolution kernels with dimensions 3×3, output features with dimensions (W/8) × (H/8) ×256;
layer 13, pooling layer, using 2×2 max pooling, stride=2, feature dimension becomes (W/16) × (H/16) ×256;
layer 14, the convolution layer, comprising 512 convolution kernels with dimensions 3×3, the feature with output dimensions (W/16) × (H/16) ×512;
layer 16, the convolution layer, contains 512 convolution kernels with dimensions 3×3, the feature with output dimensions (W/16) × (H/16) ×512;
layer 17, pooling layer, using 2×2 max pooling, stride=2, feature dimension becomes (W/32) × (H/32) ×512;
layer 18, the convolution layer, contains 512 convolution kernels with dimensions 3×3, the feature with output dimensions (W/32) × (H/32) ×512;
layer 19, the convolution layer, contains 1024 convolution kernels with dimensions of 3×3, the feature with output dimensions of (W/32) × (H/32) ×512;
and the 20 th layer is pooled, global average pooling is adopted, and the characteristic dimension is changed into 1 multiplied by 1024.
The geometric characteristic prediction module structure comprises three prediction branches, wherein each prediction branch comprises a full connection layer with the dimension of 512 and a full connection layer with the dimension of 256; the three predicted branches are a target width predicted branch, a target length predicted branch, and a target aspect ratio predicted branch.
As an optional implementation manner, in an embodiment of the present invention, the processing, by the second network model, the input slice to obtain the target geometric feature prediction value includes:
processing the input slice by utilizing the geometric feature extraction network module to obtain a feature three-dimensional vector
Figure BDA0004137981240000051
Said feature three-dimensional vector->
Figure BDA0004137981240000052
Acquiring image spatial resolution from the input slice, combining the image spatial resolution with the feature three-dimensional vector
Figure BDA0004137981240000053
Multiplying to obtain weighted geometric features;
and predicting the weighted geometric feature by utilizing the target width prediction branch to obtain a target width prediction feature
Figure BDA0004137981240000054
Said target width prediction feature->
Figure BDA0004137981240000055
Predicting the weighted geometric feature by using the target length prediction branch to obtain a target length prediction feature f l i The method comprises the steps of carrying out a first treatment on the surface of the The target length prediction feature f l i ∈d 1×1×256 ；
Predicting the weighted geometric feature by using the target aspect ratio prediction branch to obtain a target aspect ratio prediction feature
Figure BDA0004137981240000061
The target aspect ratio prediction feature->
Figure BDA0004137981240000062
Using the Relu function, for the width prediction feature
Figure BDA0004137981240000063
Target length prediction feature f l i Target aspect ratio prediction feature->
Figure BDA0004137981240000064
Processing to obtain a target predicted value; the target predicted value includes a target width predicted value +. >
Figure BDA0004137981240000065
Target Length predictor +.>
Figure BDA0004137981240000066
And target aspect ratio predictor->
Figure BDA00041379812400000620
The Relu function is:
F(w T x+b)＝max(0,w T x+b)
wherein F (·) represents the target predictor; w represents a weight matrix; max (·) represents taking the maximum value; b represents a function bias; x represents a width prediction feature
Figure BDA0004137981240000067
Or, target length prediction feature f l i Or, target aspect ratio prediction feature->
Figure BDA0004137981240000068
As an optional implementation manner, in an embodiment of the present invention, the fusion prediction model includes a fully connected layer with a dimension of 2048 and a fully connected layer with a dimension of 1024;
the fusion prediction model performs fusion processing on the target depth feature and the target geometric feature predicted value to obtain target category prediction information, and the fusion prediction model comprises the following steps:
characterizing the target depth f i With the target width predicted value
Figure BDA0004137981240000069
Target Length predictor +.>
Figure BDA00041379812400000610
And target aspect ratio predictor->
Figure BDA00041379812400000611
Sequentially fusing in series to obtain fusion characteristics +.>
Figure BDA00041379812400000612
Said fusion feature->
Figure BDA00041379812400000613
Figure BDA00041379812400000614
Utilizing the fusion prediction model to perform fusion on the fusion characteristics
Figure BDA00041379812400000615
Extracting features and reducing dimension to obtain classified features
Figure BDA00041379812400000616
Said classification feature->
Figure BDA00041379812400000617
Using a Softmax function to classify the features
Figure BDA00041379812400000618
Processing to obtain target class prediction information +.>
Figure BDA00041379812400000619
As an optional implementation manner, in an embodiment of the present invention, training the target recognition initial model fused with priori knowledge by using the training slice information set to obtain a target recognition model includes:
S41, setting the iteration times N and the maximum iteration times N max The method comprises the steps of carrying out a first treatment on the surface of the The initial value of the iteration times N is 0;
s42, constructing a loss function;
s43, inputting training slices in the training slice information set into a target identification initial model fused with priori knowledge in sequence;
s44, adjusting the parameter information of the target identification initial model fused with priori knowledge for the purpose of converging the loss function;
s45, updating the iteration times N to be N+1;
s46, judging whether the iteration number N is smaller than the maximum iteration number N max Obtaining a second judgment result;
when the second judgment result is yes, triggering and executing step S43;
and when the second judgment result is negative, finishing training, and determining the target recognition initial model fused with priori knowledge as a target recognition model.
As an alternative implementation manner, in an embodiment of the present invention, the loss function is:
L＝L F +λ(L l +L w +L r )
wherein L is F Predicting loss, L for target class l Predicting loss, L for target length w Predicting loss, L for target width r Predicting loss for target aspect ratioThe method comprises the steps of carrying out a first treatment on the surface of the λ is the weighting coefficient;
the target class prediction loss is:
Figure BDA0004137981240000071
wherein C is i The true class label for the i-th input slice,
Figure BDA0004137981240000072
for the predictive category of the ith input slice, n is the total number of input slices in a training batch;
The target length prediction loss is:
Figure BDA0004137981240000073
wherein l i A true value for the target length of the ith input slice,
Figure BDA0004137981240000081
for the predicted length of the ith input slice, n is the total number of input slices in a training batch;
the target width prediction loss is:
Figure BDA0004137981240000082
wherein w is i For the target width truth value of the ith input slice,
Figure BDA0004137981240000083
for the predicted width of the ith input slice, n is the total number of input slices in a training batch;
the target aspect ratio prediction loss L r The method comprises the following steps:
Figure BDA0004137981240000084
wherein r is i For the target aspect ratio truth value for the ith input slice,
Figure BDA0004137981240000085
for the predicted aspect ratio of the ith input slice, n is the total number of input slices in a training batch.
As an optional implementation manner, the processing the satellite remote sensing image to be detected by using the target recognition model to obtain an image target class result includes:
s51, performing noise reduction, image correction and format conversion processing on the satellite remote sensing image to be detected to obtain a first slice to be detected; the first slice to be detected is in a ". Tiff" format;
s52, respectively adjusting the width and the height of the first slice to be detected to a preset width value and a preset height value to obtain a second slice to be detected;
S53, overlapping and superposing the second target slice with a preset full-black image center point to obtain a slice to be detected; the preset full-black image represents an image with all pixel values of 0, and the width and the height of the preset full-black image are respectively the preset width value and the preset height value;
s54, performing recognition processing on the slice to be detected by using the target recognition model to obtain an image target class result.
Compared with the prior art, the embodiment of the invention has the following beneficial effects:
the invention discloses a remote sensing image target identification method integrating priori knowledge, which comprises the following steps: preprocessing the acquired satellite remote sensing image set containing the identification target to obtain a training slice information set; training the target recognition initial model fused with priori knowledge by using the training slice information set to obtain a target recognition model; and processing the satellite remote sensing image to be detected by using the target recognition model to obtain an image target class result. In the preprocessing process of the satellite remote sensing image set, the original satellite remote sensing image is marked by using priori knowledge, and the obtained marking file accelerates the training process of identifying the initial model; the established target identification initial model utilizes a mature convolutional neural network, and by adding network branches and utilizing priori knowledge, the guide model extracts geometric features such as effective target length, target width, target length-width ratio and the like, and fuses the geometric features with depth features extracted by the convolutional neural network to realize image target category identification. Therefore, the invention improves the target recognition precision of the remote sensing image and enhances the interpretability and the robustness of the target recognition model through the effective fusion of priori knowledge and machine learning.
Drawings
In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for the description of the embodiments will be briefly described below, and it is apparent that the drawings in the following description are only some embodiments of the present invention, and other drawings may be obtained according to these drawings without inventive effort for a person skilled in the art.
Fig. 1 is a schematic flow chart of a remote sensing image target recognition method with prior knowledge fusion, which is disclosed by the embodiment of the invention;
fig. 2 is a schematic diagram of an initial model structure of target recognition with prior knowledge fusion according to an embodiment of the present invention.
Detailed Description
In order to make the present invention better understood by those skilled in the art, the following description will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings, and it is apparent that the described embodiments are only some embodiments of the present invention, not all embodiments. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.
The terms first, second and the like in the description and in the claims and in the above-described figures are used for distinguishing between different objects and not necessarily for describing a sequential or chronological order. Furthermore, the terms "comprise" and "have," as well as any variations thereof, are intended to cover a non-exclusive inclusion. For example, a process, method, apparatus, article, or device that comprises a list of steps or elements is not limited to the list of steps or elements but may, in the alternative, include other steps or elements not expressly listed or inherent to such process, method, article, or device.
Reference herein to "an embodiment" means that a particular feature, structure, or characteristic described in connection with the embodiment may be included in at least one embodiment of the invention. The appearances of such phrases in various places in the specification are not necessarily all referring to the same embodiment, nor are separate or alternative embodiments mutually exclusive of other embodiments. Those of skill in the art will explicitly and implicitly appreciate that the embodiments described herein may be combined with other embodiments.
The invention discloses a remote sensing image target identification method integrating priori knowledge. The following will describe in detail.
Examples
Referring to fig. 1, fig. 1 is a schematic flow chart of a remote sensing image target recognition method with prior knowledge fusion according to an embodiment of the present invention. The remote sensing image target recognition method with the prior knowledge integrated as described in fig. 1 is applied to a satellite remote sensing image target recognition system, such as a local server or a cloud server for remote sensing image target recognition management with the prior knowledge integrated, and the embodiment of the invention is not limited. As shown in fig. 1, the remote sensing image target recognition method with fused prior knowledge may include the following operations:
s1, acquiring a satellite remote sensing image set containing an identification target;
in the embodiment of the invention, the satellite remote sensing image set comprises M satellite remote sensing images; m is a positive integer not less than 1.
S2, preprocessing the satellite remote sensing image set to obtain a training slice information set.
S3, constructing a target identification initial model fused with priori knowledge;
in the embodiment of the invention, the target recognition initial model fused with priori knowledge comprises a first network model, a second network model and a fusion prediction model.
And S4, training the target recognition initial model fused with priori knowledge by using the training slice information set to obtain a target recognition model.
S5, processing the satellite remote sensing image to be detected by using the target recognition model to obtain an image target class result;
in the embodiment of the invention, the satellite remote sensing image to be detected characterizes the satellite remote sensing image of which the target category needs to be predicted.
In this optional embodiment, as an optional implementation manner, the preprocessing of the satellite remote sensing image set in the step S2 to obtain the training slice information set includes:
s21, carrying out noise reduction, image correction and format conversion processing on satellite remote sensing images in the satellite remote sensing image set to obtain a first sample set;
in an embodiment of the present invention, the first sample includes a target slice; the target slice comprises an image slice containing a target and the spatial resolution of the image; the target slice is in ". Tiff" format.
Optionally, the noise reduction processing of the satellite remote sensing image includes gaussian noise removal, spiced salt noise removal and the like.
Optionally, performing image processing correction on the satellite remote sensing image includes geometric correction, radiation correction, and the like.
S22, labeling the first sample set according to a preset priori knowledge model to obtain a second sample set;
in the embodiment of the present invention, the second sample includes a target slice and a markup file; the annotation file comprises a target category and a target priori knowledge geometric feature; the target prior knowledge geometric features include target length, target width, and target aspect ratio. Therefore, priori knowledge is introduced in the preprocessing process of the satellite remote sensing image set, so that the method can be used for guiding intelligent identification of the target slice target and improving the identification precision of the intelligent model.
S23, filling all target slices in the second sample set to obtain a training slice information set;
in the embodiment of the invention, the training slice information includes a training slice and a labeling file.
In this optional embodiment, as an optional implementation manner, the filling processing is performed on all target slices in the second sample set to obtain a training slice information set, including:
s231, any target slice in the second sample set is acquired;
s232, respectively adjusting the width and the height of any target slice to a preset width value and a preset height value to obtain target slices with the same width and height; the preset width value is not smaller than the width of the target slice; the preset height value is not smaller than the height of the target slice;
s233, overlapping and superposing the target slices with the same width and height with a preset full-black image center point to obtain a training slice; the preset full-black image represents an image with all pixel values of 0, and the width and the height of the preset full-black image are respectively the preset width value and the preset height value;
s234, updating any target slice in the second sample set to be a training slice;
s235, judging whether all target slices in the second sample set are updated to training slices, and obtaining a first judgment result;
If the first judgment result is negative, updating the target slice which is not updated into any target slice, and triggering and executing the step S232;
and if the first judgment result is yes, the updated second sample set is a training slice information set.
Therefore, by implementing the remote sensing image target recognition method with the prior knowledge fusion described in the embodiment of the invention, the target recognition initial model recognition training is guided by introducing the prior knowledge accumulated by artificial visual interpretation in the pretreatment of the satellite remote sensing image set and the training process of the target recognition model, so that the effective fusion of the prior knowledge and the machine learning is realized, the model recognition precision is improved, and the model interpretability and the robustness are enhanced.
In an alternative embodiment, as shown in fig. 2, fig. 2 is a schematic structural diagram of an object recognition initial model fused with priori knowledge according to an embodiment of the present invention, where the constructing an object recognition initial model fused with priori knowledge in the step S3 includes:
s31, constructing a first network model; the first network model is used for processing the input slice to obtain a target depth characteristic f i The method comprises the steps of carrying out a first treatment on the surface of the The target depth feature f i ∈d 1×1×2048 ；
S32, constructing a second network model; the second network model is used for processing the input slice to obtain a target geometric feature predicted value; the target geometric characteristic predicted value comprises a target width predicted value, a target length predicted value and a target length-width ratio predicted value;
S33, constructing a fusion prediction model; and the fusion prediction model is used for carrying out fusion processing on the target depth feature and the target geometric feature predicted value to obtain target category prediction information.
In this alternative embodiment, as an alternative implementation manner, the first network model is a convolutional neural network VGG, or, res net, or, seNet, or, shuffleNet, or, googleNet.
In this optional embodiment, as an optional implementation manner, the second network model includes a geometric feature extraction network module and a geometric feature prediction module.
The geometric feature extraction network module includes:
layer 1, input layer, dimension W×H×N; w, H and N respectively represent the width, height and channel number of an input slice;
in this embodiment, W is the preset width value, and H is the preset height value;
layer 2, the convolution layer, comprising 64 convolution kernels of dimension 3 x 3, outputting features of dimension W x H x 64;
layer 3, the convolution layer, comprising 64 convolution kernels of dimension 3×3, outputting features of dimension w×h×64;
layer 4, pooling layer, using 2×2 max pooling, stride=2, feature dimension becomes (W/2) × (H/2) ×64;
Layer 5, the convolution layer, comprising 128 convolution kernels of dimension 3 x 3, outputting features of dimension (W/2) x (H/2) x 128;
layer 6, the convolution layer, comprising 128 convolution kernels of dimension 3 x 3, outputting features of dimension (W/2) x (H/2) x 128;
layer 7, pooling layer, using 2×2 max pooling, stride=2, feature dimension becomes (W/4) × (H/4) ×128;
layer 8, the convolution layer, contains 256 convolution kernels with dimensions of 3×3, output features with dimensions of (W/4) × (H/4) ×256;
layer 9, the convolution layer, contains 256 convolution kernels with dimensions of 3×3, output features with dimensions of (W/4) × (H/4) ×256;
layer 10, pooling layer, using 2×2 max pooling, stride=2, feature dimension becomes (W/8) × (H/8) ×256;
layer 11, the convolution layer, contains 256 convolution kernels with dimensions of 3 x 3, output features with dimensions of (W/8) × (H/8) ×256;
layer 12, the convolution layer, contains 256 convolution kernels with dimensions 3×3, output features with dimensions (W/8) × (H/8) ×256;
layer 13, pooling layer, using 2×2 max pooling, stride=2, feature dimension becomes (W/16) × (H/16) ×256;
layer 14, the convolution layer, comprising 512 convolution kernels with dimensions 3×3, the feature with output dimensions (W/16) × (H/16) ×512;
Layer 16, the convolution layer, contains 512 convolution kernels with dimensions 3×3, the feature with output dimensions (W/16) × (H/16) ×512;
layer 17, pooling layer, using 2×2 max pooling, stride=2, feature dimension becomes (W/32) × (H/32) ×512;
layer 18, the convolution layer, contains 512 convolution kernels with dimensions 3×3, the feature with output dimensions (W/32) × (H/32) ×512;
layer 19, the convolution layer, contains 1024 convolution kernels with dimensions of 3×3, the feature with output dimensions of (W/32) × (H/32) ×512;
and the 20 th layer is pooled, global average pooling is adopted, and the characteristic dimension is changed into 1 multiplied by 1024.
It can be seen that by using the above-mentioned geometric feature extraction network module, the beneficial effect of extracting the target geometric feature can be achieved.
The geometric characteristic prediction module structure comprises three prediction branches, wherein each prediction branch comprises a full connection layer with the dimension of 512 and a full connection layer with the dimension of 256; the three predicted branches are a target width predicted branch, a target length predicted branch, and a target aspect ratio predicted branch.
Therefore, three prediction branches are utilized to extract target geometric features, and important identification features used in the process of the interpretation of an interpreter are integrated into intelligent identification, so that the interpretation experience of a person is integrated into an intelligent model.
As an optional implementation manner, in an embodiment of the present invention, the processing, by the second network model, the input slice to obtain the target geometric feature prediction value includes:
processing the input slice by using the geometric feature extraction network module to obtain a feature three-dimensional vector
Figure BDA0004137981240000151
The characteristic three-dimensional vector->
Figure BDA0004137981240000152
Acquiring an image spatial resolution from the input slice, and combining the image spatial resolution with the feature three-dimensional vector
Figure BDA0004137981240000153
Multiplying to obtain weighted geometric features;
predicting the weighted geometric feature by using the target width prediction branch to obtain a target width prediction feature
Figure BDA0004137981240000154
The target width prediction featureSyndrome of->
Figure BDA0004137981240000155
Predicting the weighted geometric feature by using the target length prediction branch to obtain a target length prediction feature f l i The method comprises the steps of carrying out a first treatment on the surface of the The target length prediction feature f l i ∈d 1×1×256 ；
Predicting the weighted geometric feature by using the target aspect ratio prediction branch to obtain a target aspect ratio prediction feature
Figure BDA0004137981240000156
The above target aspect ratio prediction feature->
Figure BDA0004137981240000157
For the width prediction feature using the Relu function
Figure BDA0004137981240000158
Target length prediction feature f l i Target aspect ratio prediction feature->
Figure BDA0004137981240000159
Processing to obtain a target predicted value; the target predicted value includes a target width predicted value +. >
Figure BDA00041379812400001510
Target Length predictor +.>
Figure BDA00041379812400001511
And target aspect ratio predictor->
Figure BDA00041379812400001512
The Relu function described above is:
F(w T x+b)＝max(0,w T x+b)
wherein F (·) represents the target predictor;w represents a weight matrix; max (·) represents taking the maximum value; b represents a function bias; x represents a width prediction feature
Figure BDA0004137981240000161
Or, target length prediction feature f l i Or, target aspect ratio prediction feature->
Figure BDA0004137981240000162
Therefore, the target geometric feature predicted value can be calculated by using the second network model, and the guiding model extracts the target geometric feature.
As an optional implementation manner, in an embodiment of the present invention, the fusion prediction model includes a fully connected layer with a dimension of 2048 and a fully connected layer with a dimension of 1024;
the fusion prediction model performs fusion processing on the target depth feature and the target geometric feature predicted value to obtain target category prediction information, and the fusion prediction model comprises the following steps:
the target depth feature f i The target width predicted value is the same as the target width predicted value
Figure BDA0004137981240000163
Target Length predictor +.>
Figure BDA0004137981240000164
And target aspect ratio predictor->
Figure BDA0004137981240000165
Sequentially fusing in series to obtain fusion characteristics +.>
Figure BDA0004137981240000166
Fusion characteristics described above->
Figure BDA0004137981240000167
Figure BDA0004137981240000168
Using the fusion prediction model to perform fusion feature
Figure BDA0004137981240000169
Extracting features and reducing dimensions to obtain 1024-dimensional classification features +.>
Figure BDA00041379812400001610
The above classification feature->
Figure BDA00041379812400001611
The above classification features are characterized by using a Softmax function
Figure BDA00041379812400001612
Processing to obtain target class prediction information +.>
Figure BDA00041379812400001613
As an optional implementation manner, in an embodiment of the present invention, training the target recognition initial model fused with priori knowledge by using the training slice information set to obtain a target recognition model includes:
s41, setting the iteration times N and the maximum iteration times N max The method comprises the steps of carrying out a first treatment on the surface of the The initial value of the iteration number N is 0;
s42, constructing a loss function;
s43, inputting the training slices in the training slice information set into a target identification initial model fused with priori knowledge in sequence;
s44, adjusting the parameter information of the target recognition initial model fused with priori knowledge for the purpose of converging the loss function;
s45, updating iteration times N to be N+1;
s46, judging whether the iteration number N is smaller than the maximum iteration number N max Obtaining a second judgment result;
when the second judgment result is yes, triggering and executing the step S41;
and when the second judgment result is negative, finishing training, and determining the target recognition initial model fused with the priori knowledge as a target recognition model.
As an alternative implementation manner, in the embodiment of the present invention, the loss function is:
L＝L F +λ(L l +L w +L r )
wherein L is F Predicting loss, L for target class l Predicting loss, L for target length w Predicting loss, L for target width r Predicting loss for a target aspect ratio; λ is the weighting coefficient.
Preferably, λ=0.5.
The target class prediction loss is:
Figure BDA0004137981240000171
wherein C is i Category labels in the markup file for the ith input slice,
Figure BDA0004137981240000172
for the predictive category of the ith input slice, n is the total number of input slices in a training batch;
the target length prediction loss is as follows:
Figure BDA0004137981240000173
wherein l i A true value for the target length of the ith input slice,
Figure BDA0004137981240000174
for the predicted length of the ith input slice, n is the total number of input slices in a training batch;
the target width prediction loss is:
Figure BDA0004137981240000175
wherein w is i For the target width truth value of the ith input slice,
Figure BDA0004137981240000176
for the predicted width of the ith input slice, n is the total number of input slices in a training batch;
the target aspect ratio prediction loss L r The method comprises the following steps:
Figure BDA0004137981240000181
wherein r is i For the target aspect ratio truth value for the ith input slice,
Figure BDA0004137981240000182
for the predicted aspect ratio of the ith input slice, n is the total number of input slices in a training batch.
As an optional implementation manner, in the step S5, the processing the satellite remote sensing image to be detected by using the target recognition model to obtain the image target class result includes:
S51, carrying out noise reduction, image correction and format conversion processing on the satellite remote sensing image to be detected to obtain a first slice to be detected; the first slice to be measured is in a ". Tiff" format;
s52, respectively adjusting the width and the height of the first slice to be detected to a preset width value and a preset height value to obtain a second slice to be detected;
s53, overlapping and superposing a second target slice with a preset full-black image center point to obtain a slice to be detected; the preset full-black image represents an image with all pixel values of 0, and the width and the height of the preset full-black image are respectively the preset width value and the preset height value;
s54, identifying the slice to be detected by using the target identification model to obtain an image target class result.
From the above detailed description of the embodiments, it will be apparent to those skilled in the art that the embodiments may be implemented by means of software plus necessary general hardware platforms, or of course by means of hardware. Based on such understanding, the foregoing technical solutions may be embodied essentially or in part in the form of a software product that may be stored in a computer-readable storage medium including Read-Only Memory (ROM), random-access Memory (Random Access Memory, RAM), programmable Read-Only Memory (Programmable Read-Only Memory, PROM), erasable programmable Read-Only Memory (Erasable Programmable Read Only Memory, EPROM), one-time programmable Read-Only Memory (OTPROM), electrically erasable programmable Read-Only Memory (EEPROM), compact disc Read-Only Memory (Compact Disc Read-Only Memory, CD-ROM) or other optical disc Memory, magnetic disc Memory, tape Memory, or any other medium that can be used for computer-readable carrying or storing data.
Finally, it should be noted that: the remote sensing image target recognition method with the prior knowledge disclosed by the embodiment of the invention is disclosed as a preferred embodiment of the invention, and is only used for illustrating the technical scheme of the invention, but not limiting the technical scheme; although the invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art will understand that; the technical scheme recorded in the various embodiments can be modified or part of technical features in the technical scheme can be replaced equivalently; such modifications and substitutions do not depart from the spirit and scope of the corresponding technical solutions.

Claims (9)

1. The remote sensing image target recognition method integrating priori knowledge is characterized by comprising the following steps of:
s1, acquiring a satellite remote sensing image set containing an identification target; the satellite remote sensing image set comprises M satellite remote sensing images; m is a positive integer not less than 1;
s2, preprocessing the satellite remote sensing image set to obtain a training slice information set;
s3, constructing a target identification initial model fused with priori knowledge; the target identification initial model integrating priori knowledge comprises a first network model, a second network model and an integrated prediction model;
S4, training the target recognition initial model fused with priori knowledge by utilizing the training slice information set to obtain a target recognition model;
s5, processing the satellite remote sensing image to be detected by using the target recognition model to obtain an image target class result; the satellite remote sensing image to be detected characterizes a satellite remote sensing image of which the target category needs to be predicted.
2. The method for identifying a target of a remote sensing image fused with priori knowledge according to claim 1, wherein the preprocessing the satellite remote sensing image set to obtain a training slice information set comprises:
s21, carrying out noise reduction, image correction and format conversion processing on the satellite remote sensing image set to obtain a first sample set; the first sample includes a target slice; the target slice comprises an image slice containing a target and image space resolution; the target slice is in a ". Tiff" format;
s22, labeling the first sample set according to a preset priori knowledge model to obtain a second sample set; the second sample comprises a target slice and a labeling file; the annotation file comprises a target category and a target priori knowledge geometric feature; the target priori knowledge geometric features comprise a target length, a target width and a target aspect ratio;
S23, filling all target slices in the second sample set to obtain a training slice information set; the training slice information comprises a training slice and a labeling file.
3. The method for identifying a target of a remote sensing image with integrated prior knowledge according to claim 2, wherein the filling all target slices in the second sample set to obtain a training slice information set includes:
s231, any target slice in the second sample set is acquired;
s232, respectively adjusting the width and the height of any target slice to a preset width value and a preset height value to obtain target slices with the same width and height; the preset width value is not smaller than the width of the target slice; the preset height value is not smaller than the height of the target slice;
s233, overlapping and superposing the target slices with the same width and height with a preset full-black image center point to obtain a training slice; the preset full-black image represents an image with all pixel values of 0, and the width and the height of the preset full-black image are respectively the preset width value and the preset height value;
s234, updating any target slice in the second sample set to be a training slice;
S235, judging whether all target slices in the second sample set are updated to training slices, and obtaining a first judgment result;
if the first judgment result is negative, updating the target slice which is not updated into any target slice, and triggering and executing the step S232;
and if the first judgment result is yes, the updated second sample set is a training slice information set.
4. The method for identifying a target of a remote sensing image fused with priori knowledge according to claim 1, wherein the constructing an initial model of target identification fused with priori knowledge comprises:
s31, constructing a first network model; the first network model is used for processing an input slice to obtain a target depth characteristic f i The method comprises the steps of carrying out a first treatment on the surface of the The target depth feature f i ∈d 1×1×2048 ；
S32, constructing a second network model; the second network model is used for processing the input slice to obtain a target geometric characteristic predicted value; the target geometric characteristic predicted value comprises a target width predicted value, a target length predicted value and a target length-width ratio predicted value;
s33, constructing a fusion prediction model; and the fusion prediction model is used for carrying out fusion processing on the target depth characteristic and the target geometric characteristic predicted value to obtain target category predicted information.
5. The method for identifying a target of a remote sensing image with integrated prior knowledge according to claim 4, wherein the second network model comprises a geometric feature extraction network module and a geometric feature prediction module;
the geometric feature extraction network module comprises:
layer 1, input layer, dimension W×H×N; w, H and N are the width, height and channel number of the input slice respectively;
layer 2, the convolution layer, comprising 64 convolution kernels of dimension 3 x 3, outputting features of dimension W x H x 64;
layer 3, the convolution layer, comprising 64 convolution kernels of dimension 3×3, outputting features of dimension w×h×64;
layer 4, pooling layer, using 2×2 max pooling, stride=2, feature dimension becomes (W/2) × (H/2) ×64;
layer 5, the convolution layer, comprising 128 convolution kernels of dimension 3 x 3, outputting features of dimension (W/2) x (H/2) x 128;
layer 6, the convolution layer, comprising 128 convolution kernels of dimension 3 x 3, outputting features of dimension (W/2) x (H/2) x 128;
layer 7, pooling layer, using 2×2 max pooling, stride=2, feature dimension becomes (W/4) × (H/4) ×128;
layer 8, the convolution layer, contains 256 convolution kernels with dimensions of 3×3, output features with dimensions of (W/4) × (H/4) ×256;
Layer 9, the convolution layer, contains 256 convolution kernels with dimensions of 3×3, output features with dimensions of (W/4) × (H/4) ×256;
layer 10, pooling layer, using 2×2 max pooling, stride=2, feature dimension becomes (W/8) × (H/8) ×256;
layer 11, the convolution layer, contains 256 convolution kernels with dimensions of 3 x 3, output features with dimensions of (W/8) × (H/8) ×256;
layer 12, the convolution layer, contains 256 convolution kernels with dimensions 3×3, output features with dimensions (W/8) × (H/8) ×256;
layer 13, pooling layer, using 2×2 max pooling, stride=2, feature dimension becomes (W/16) × (H/16) ×256;
layer 14, the convolution layer, comprising 512 convolution kernels with dimensions 3×3, the feature with output dimensions (W/16) × (H/16) ×512;
layer 16, the convolution layer, contains 512 convolution kernels with dimensions 3×3, the feature with output dimensions (W/16) × (H/16) ×512;
layer 17, pooling layer, using 2×2 max pooling, stride=2, feature dimension becomes (W/32) × (H/32) ×512;
layer 18, the convolution layer, contains 512 convolution kernels with dimensions 3×3, the feature with output dimensions (W/32) × (H/32) ×512;
layer 19, the convolution layer, contains 1024 convolution kernels with dimensions of 3×3, the feature with output dimensions of (W/32) × (H/32) ×512;
The 20 th layer is pooled, global average pooling is adopted, and the characteristic dimension is changed into 1 multiplied by 1024;
the geometric characteristic prediction module structure comprises three prediction branches, wherein each prediction branch comprises a full connection layer with the dimension of 512 and a full connection layer with the dimension of 256; the three predicted branches are a target width predicted branch, a target length predicted branch, and a target aspect ratio predicted branch.
6. The method for identifying a target of a remote sensing image with integrated prior knowledge according to claim 5, wherein the second network model is used for processing an input slice to obtain a target geometric feature predicted value, and the method comprises the following steps:
processing the input slice by utilizing the geometric feature extraction network module to obtain a feature three-dimensional vector
Figure FDA0004137981230000041
Said feature three-dimensional vector->
Figure FDA0004137981230000042
Acquiring image spatial resolution from the input slice, combining the image spatial resolution with the feature three-dimensional vector
Figure FDA0004137981230000043
Multiplying to obtain weighted geometric features;
and predicting the weighted geometric feature by utilizing the target width prediction branch to obtain a target width prediction feature
Figure FDA0004137981230000051
Said target width prediction feature->
Figure FDA0004137981230000052
Predicting the weighted geometric feature by using the target length prediction branch to obtain a target length prediction feature f l i The method comprises the steps of carrying out a first treatment on the surface of the The target length prediction feature
Figure FDA0004137981230000053
Predicting the weighted geometric feature by using the target aspect ratio prediction branch to obtain a target aspect ratio prediction feature
Figure FDA0004137981230000054
The target aspect ratio prediction feature->
Figure FDA0004137981230000055
Using the Relu function, for the width prediction feature
Figure FDA0004137981230000056
Target length prediction feature f l i Target aspect ratio prediction feature->
Figure FDA00041379812300000521
Processing to obtain a target predicted value; the target predicted value includes a target width predicted value +.>
Figure FDA0004137981230000057
Target Length predictor +.>
Figure FDA0004137981230000058
And target aspect ratio predictor->
Figure FDA0004137981230000059
The Relu function is:
F(w T x+b)＝max(0,w T x+b)
wherein F (·) represents the target predictor; w represents a weight matrix; max (·) represents taking the maximum value; b represents a function bias; x represents a width prediction feature
Figure FDA00041379812300000510
Or, target length prediction feature f l i Or, target aspect ratio prediction feature->
Figure FDA00041379812300000511
7. The method for identifying a remote sensing image target by fusing priori knowledge according to claim 6, wherein the fused prediction model comprises a fully connected layer with a dimension of 2048 and a fully connected layer with a dimension of 1024;
the fusion prediction model performs fusion processing on the target depth feature and the target geometric feature predicted value to obtain target category prediction information, and the fusion prediction model comprises the following steps:
characterizing the target depth f i With the target width predicted value
Figure FDA00041379812300000512
Target Length predictor +.>
Figure FDA00041379812300000513
And target aspect ratio predictor->
Figure FDA00041379812300000522
Sequentially fusing in series to obtain fusion characteristics +.>
Figure FDA00041379812300000514
Said fusion feature->
Figure FDA00041379812300000515
Figure FDA00041379812300000516
Utilizing the fusion prediction model to perform fusion on the fusion characteristics
Figure FDA00041379812300000517
Extracting features and reducing dimension to obtain classified features ∈K>
Figure FDA00041379812300000518
Said classification feature->
Figure FDA00041379812300000519
Using a Softmax function to classify the features
Figure FDA00041379812300000520
Processing to obtain target class prediction information +.>
Figure FDA0004137981230000061
8. The method for identifying a target of a remote sensing image with integrated priori knowledge according to claim 7, wherein training the initial model of target identification with integrated priori knowledge by using the training slice information set to obtain the model of target identification comprises:
s41, setting the iteration times N and the maximum iteration times N max The method comprises the steps of carrying out a first treatment on the surface of the The initial value of the iteration times N is 0;
s42, constructing a loss function;
s43, inputting training slices in the training slice information set into a target identification initial model fused with priori knowledge in sequence;
s44, adjusting the parameter information of the target identification initial model fused with priori knowledge for the purpose of converging the loss function;
s45, updating the iteration times N to be N+1;
s46, judging whether the iteration number N is smaller than the maximum iteration number N max Obtaining a second judgment result;
when the second judgment result is yes, triggering and executing step S43;
and when the second judgment result is negative, finishing training, and determining the target recognition initial model fused with priori knowledge as a target recognition model.
9. The method for identifying a target of a remote sensing image fused with priori knowledge according to claim 8, wherein the processing the satellite remote sensing image to be detected by using the target identification model to obtain an image target class result comprises:
s51, performing noise reduction, image correction and format conversion processing on the satellite remote sensing image to be detected to obtain a first slice to be detected; the first slice to be detected is in a ". Tiff" format;
s52, respectively adjusting the width and the height of the first slice to be detected to a preset width value and a preset height value to obtain a second slice to be detected;
s53, overlapping and superposing the second target slice with a preset full-black image center point to obtain a slice to be detected; the preset full-black image represents an image with all pixel values of 0, and the width and the height of the preset full-black image are respectively the preset width value and the preset height value;
S54, performing recognition processing on the slice to be detected by using the target recognition model to obtain an image target class result.
CN202310280652.4A 2023-03-21 2023-03-21 Remote sensing image target recognition method integrating priori knowledge Pending CN116434076A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202310280652.4A CN116434076A (en) 2023-03-21 2023-03-21 Remote sensing image target recognition method integrating priori knowledge

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202310280652.4A CN116434076A (en) 2023-03-21 2023-03-21 Remote sensing image target recognition method integrating priori knowledge

Publications (1)

Publication Number Publication Date
CN116434076A true CN116434076A (en) 2023-07-14

Family

ID=87084680

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202310280652.4A Pending CN116434076A (en) 2023-03-21 2023-03-21 Remote sensing image target recognition method integrating priori knowledge

Country Status (1)

Country Link
CN (1) CN116434076A (en)

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112084941A (en) * 2020-09-09 2020-12-15 国科天成（北京）科技有限公司 Target detection and identification method based on remote sensing image
US20220188366A1 (en) * 2020-12-15 2022-06-16 NantMedia Holdings, LLC Combined Wide And Deep Machine Learning Models For Automated Database Element Processing Systems, Methods And Apparatuses
CN115457396A (en) * 2022-09-26 2022-12-09 河北省科学院地理科学研究所 Surface target ground object detection method based on remote sensing image
CN115496752A (en) * 2022-11-16 2022-12-20 齐鲁工业大学 Steel surface defect detection method based on one-stage target detection algorithm
CN115588140A (en) * 2022-10-24 2023-01-10 北京市遥感信息研究所 Multi-spectral remote sensing image multi-directional target detection method

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112084941A (en) * 2020-09-09 2020-12-15 国科天成（北京）科技有限公司 Target detection and identification method based on remote sensing image
US20220188366A1 (en) * 2020-12-15 2022-06-16 NantMedia Holdings, LLC Combined Wide And Deep Machine Learning Models For Automated Database Element Processing Systems, Methods And Apparatuses
CN115457396A (en) * 2022-09-26 2022-12-09 河北省科学院地理科学研究所 Surface target ground object detection method based on remote sensing image
CN115588140A (en) * 2022-10-24 2023-01-10 北京市遥感信息研究所 Multi-spectral remote sensing image multi-directional target detection method
CN115496752A (en) * 2022-11-16 2022-12-20 齐鲁工业大学 Steel surface defect detection method based on one-stage target detection algorithm

Similar Documents

Publication Publication Date Title
Zhang et al. Remote sensing image spatiotemporal fusion using a generative adversarial network
Xie et al. Multilevel cloud detection in remote sensing images based on deep learning
CN111931684B (en) Weak and small target detection method based on video satellite data identification features
CN106909924B (en) Remote sensing image rapid retrieval method based on depth significance
Sirmacek et al. Urban-area and building detection using SIFT keypoints and graph theory
CN111126258A (en) Image recognition method and related device
CN107918776B (en) Land planning method and system based on machine vision and electronic equipment
CN111680176A (en) Remote sensing image retrieval method and system based on attention and bidirectional feature fusion
CN112541508A (en) Fruit segmentation and recognition method and system and fruit picking robot
CN113096138B (en) Weak supervision semantic image segmentation method for selective pixel affinity learning
CN112232371B (en) American license plate recognition method based on YOLOv3 and text recognition
CN113609889B (en) High-resolution remote sensing image vegetation extraction method based on sensitive characteristic focusing perception
CN113033520A (en) Tree nematode disease wood identification method and system based on deep learning
CN113111716B (en) Remote sensing image semiautomatic labeling method and device based on deep learning
CN116524369B (en) Remote sensing image segmentation model construction method and device and remote sensing image interpretation method
CN114037640A (en) Image generation method and device
CN117274768A (en) Training method of target detection network, target detection method and related device
Carballal et al. Transfer learning features for predicting aesthetics through a novel hybrid machine learning method
CN115311508A (en) Single-frame image infrared dim target detection method based on depth U-type network
CN113516194B (en) Semi-supervised classification method, device, equipment and storage medium for hyperspectral remote sensing images
CN113762251A (en) Target classification method and system based on attention mechanism
CN117636298A (en) Vehicle re-identification method, system and storage medium based on multi-scale feature learning
CN114596515A (en) Target object detection method and device, electronic equipment and storage medium
CN116758411A (en) Ship small target detection method based on remote sensing image pixel-by-pixel processing
CN116434076A (en) Remote sensing image target recognition method integrating priori knowledge

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination