CN115858846B - Skier image retrieval method and system based on deep learning - Google Patents
Skier image retrieval method and system based on deep learning Download PDFInfo
- Publication number
- CN115858846B CN115858846B CN202310120051.7A CN202310120051A CN115858846B CN 115858846 B CN115858846 B CN 115858846B CN 202310120051 A CN202310120051 A CN 202310120051A CN 115858846 B CN115858846 B CN 115858846B
- Authority
- CN
- China
- Prior art keywords
- image
- target
- image set
- retrieval
- identification
- 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.)
- Active
Links
Images
Classifications
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
- Y02D10/00—Energy efficient computing, e.g. low power processors, power management or thermal management
Landscapes
- Image Analysis (AREA)
Abstract
The invention provides a skier image retrieval method and system based on deep learning, which relate to the technical field of image retrieval, and are used for constructing a target detection image set and training a target detection model; configuring an image preprocessing space to perform image preprocessing, and generating a re-identification image set; the Object-reID network frame based training is used for generating a feature extraction model, an image feature processing module is integrated and generated, retrieval and identification are carried out based on a retrieval request, target features are obtained, a storage database is traversed to determine an image retrieval result, the problems that in the prior art, the image feature extraction is single and the matching degree is insufficient, a certain deviation exists in image retrieval accuracy, the technical problems of missed detection and false detection exist are solved, a plurality of functional models are embedded by constructing the image feature processing module, the completeness of the image features is efficiently and accurately identified, and the high-matching-degree image retrieval is completed are solved.
Description
Technical Field
The invention relates to the technical field of image retrieval, in particular to a skier image retrieval method and system based on deep learning.
Background
The popularization of electronic mobile equipment causes mass image generation at any time, and image retrieval technology is generated, so that image retrieval becomes a main problem to be solved. For example, a large number of ski images are captured, and how to efficiently and accurately search for a desired image is the current solving direction. In order to facilitate relevant image retrieval of different skiers, currently, along with updating and iteration of an image retrieval technology, image retrieval is mainly performed by technologies such as image searching, pedestrian re-recognition and the like, and compared with a traditional retrieval method, the retrieval effect is improved to a certain extent in the retrieval efficiency direction, but certain defects still exist, the expected retrieval requirement cannot be met, and further technical innovation is needed.
In the prior art, the intelligent degree and the strict degree of a skier image retrieval method are insufficient, so that the image features are extracted more singly and the matching degree is insufficient, and the image retrieval accuracy is biased to a certain degree, so that missed detection and false detection exist.
Disclosure of Invention
The application provides a skier image retrieval method and system based on deep learning, which are used for solving the technical problems of omission and false detection caused by certain deviation of image retrieval accuracy due to single extraction of image features and insufficient matching degree caused by insufficient intelligence and rigor of the skier image retrieval method in the prior art.
In view of the above, the present application provides a skier image retrieval method and system based on deep learning.
In a first aspect, the present application provides a skier image retrieval method based on deep learning, the method comprising:
connecting the Internet of things, taking a multidimensional scene as an index direction to acquire a skiing image set, wherein the skiing image set comprises an image storage position and a storage path, and constructing a target detection image set;
extracting an identification positioning image set based on the target detection image set, and training a target detection model, wherein the identification positioning image set is provided with a skier position frame selection mark;
configuring an image preprocessing space, and performing image preprocessing on the identification positioning image set based on the image preprocessing space to generate a re-identification image set;
generating a feature extraction model based on Object-reID network framework training based on the re-identified image set;
constructing an image feature processing module based on the target detection model, the image preprocessing space and the feature extraction model;
acquiring a retrieval request of a target user, and carrying out retrieval identification on a target image according to the image feature processing module to acquire target features;
traversing a storage database, and determining an image retrieval result based on the target feature, wherein the image retrieval result is expressed as a retrieval image-storage position-storage path, and the storage position and the storage path are additional output information.
In a second aspect, the present application provides a deep learning based skier image retrieval system, the system comprising:
the image set construction module is used for connecting the Internet of things, taking the multidimensional scene as an index direction to acquire a skiing image set, and constructing a target detection image set by comprising an image storage position and a storage path;
the target detection model training module is used for extracting an identification positioning image set based on the target detection image set and training a target detection model, wherein the identification positioning image set is provided with a skier position frame selection mark;
the image preprocessing module is used for configuring an image preprocessing space, and performing image preprocessing on the identification positioning image set based on the image preprocessing space to generate a re-identification image set;
the feature extraction model training module is used for generating a feature extraction model based on the re-identification image set and Object-reID network frame training;
the building module is used for building an image feature processing module based on the target detection model, the image preprocessing space and the feature extraction model;
the feature acquisition module is used for acquiring a retrieval request of a target user, and carrying out retrieval identification on a target image according to the image feature processing module to acquire target features;
and the result determining module is used for traversing a storage database and determining an image retrieval result based on the target feature, wherein the image retrieval result is expressed as a retrieval image-storage position-storage path, and the storage position and the storage path are additional output information.
One or more technical solutions provided in the present application have at least the following technical effects or advantages:
according to the skier image retrieval method based on deep learning, which is provided by the embodiment of the application, the internet of things is connected, a skier image set is acquired by taking a multi-dimensional scene as an index direction, a target detection image set is constructed, an identification positioning image set is extracted to train a target detection model, and the identification positioning image set is provided with a skier position frame selection mark; configuring an image preprocessing space, performing image preprocessing on the identification positioning image set, and generating a re-identification image set; generating a feature extraction model based on Object-reID network framework training based on the re-identified image set; integrating and regulating the target detection model, the image preprocessing space and the feature extraction model to construct an image feature processing module; acquiring a retrieval request of a target user, and carrying out retrieval identification on a target image based on the image feature processing module to acquire target features; traversing a storage database, determining an image retrieval result based on the target features, solving the technical problems that in the prior art, the retrieval method for skier images is insufficient in intelligence and accuracy, so that the extraction of the image features is single and the matching degree is insufficient, the image retrieval accuracy has certain deviation, missed detection and false detection are caused, and the image feature processing module is built to embed a plurality of functional models, thereby realizing the complete efficient and accurate identification of the image features and completing the high-matching-degree image retrieval.
Drawings
FIG. 1 is a schematic flow chart of a skier image retrieval method based on deep learning;
FIG. 2 is a schematic diagram of a re-identification image set acquisition process in a skier image retrieval method based on deep learning;
FIG. 3 is a schematic diagram of an image retrieval result obtaining process in a skier image retrieval method based on deep learning;
fig. 4 is a schematic structural diagram of a skier image retrieval system based on deep learning.
Reference numerals illustrate: the device comprises an image set constructing module 11, a target detection model training module 12, an image preprocessing module 13, a feature extraction model training module 14, a constructing module 15, a feature acquisition module 16 and a result determining module 17.
Detailed Description
The skier image retrieval method and system based on deep learning are provided, a target detection image set is constructed, and a training target detection model of the identification positioning image set is extracted; configuring an image preprocessing space to perform image preprocessing, and generating a re-identification image set; the Object-reID network framework based training generation feature extraction model is integrated with a regular generation image feature processing module, search and identification are carried out based on a search request, target features are obtained, and a storage database is traversed to determine an image search result, so that the technical problems of missing detection and false detection caused by certain deviation in image search accuracy due to single extraction of image features and insufficient matching degree caused by insufficient intelligence and rigor of a skier image search method in the prior art are solved.
Example 1
As shown in fig. 1, the present application provides a skier image retrieval method based on deep learning, the method comprising:
step S100: connecting the Internet of things, taking a multidimensional scene as an index direction to acquire a skiing image set, wherein the skiing image set comprises an image storage position and a storage path, and constructing a target detection image set;
specifically, mass image generation exists at any time due to popularization of electronic mobile equipment, an image retrieval technology is developed, a large number of skiing images are shot, and in order to facilitate relevant image retrieval of different skiers and improve image retrieval accuracy and retrieval efficiency, the application provides a skier image retrieval method based on deep learning, modeling is performed based on a deep learning algorithm, a model integration image feature recognition extraction mechanism is performed, character fitness of recognition features can be ensured, and retrieval determination of image retrieval results is further performed in a storage database. Specifically, the method comprises the steps of connecting the internet of things, determining a multi-dimensional acquisition scene comprising various weather conditions, indoor and outdoor, time axes and the like, acquiring skiing images comprising a plurality of skiing images of different target persons, determining storage positions and storage paths of the images, identifying the acquired images and generating the target detection image set. The acquisition of the target detection image set provides basic source data for subsequent model training.
Further, step S100 of the present application further includes:
step S110: connecting with the Internet, and collecting a skiing image set in a multi-dimensional scene as a target image set;
step S120: performing multistage light color adjustment and receptive field adjustment on the target image set to determine an augmented image set;
step S130: and integrating and regulating the augmentation image set and the target image set to generate the target detection image set.
Specifically, the internet is an auxiliary system tool for performing skiing image acquisition, the internet is connected, a multi-dimensional acquisition scene such as sunny day, snowy day, foggy day, night, outdoor and the like is determined, the multi-dimensional acquisition scene is used as an index direction of skiing images, associated image acquisition is performed, the acquired images are classified and integrated based on target characters, the target characters are skiers in the images, the target image set is acquired, and the target image set comprises multiple skiing types. Setting a multistage light color adjustment scale and a receptive field adjustment scale, for example, adjusting the direction of a light source, brightness, color gamut, target person field of view, and the like, respectively adjusting the target image set for multiple times, and taking an adjusted image as the augmented image set. The augmented image set is further added to the target image set to generate the target detection image set. By image augmentation, the generalization degree of the image can be effectively improved, and the completeness of the image information can be improved.
Step S200: extracting an identification positioning image set based on the target detection image set, and training a target detection model, wherein the identification positioning image set is provided with a skier position frame selection mark;
further, the step S200 of training the target detection model based on the target detection image set to extract the identification positioning image set further includes:
step S210: extracting the identification positioning image set based on the target detection image set;
step S220: taking the target detection image set as level identification information, and taking the identification positioning image set as level decision information to construct a sample data set, wherein the target detection image set corresponds to the identification positioning image set one by one;
step S230: dividing the sample data set into K groups, and obtaining K groups of sample data subsets;
step S240: and dividing the K groups of sample data subsets into a training set and a testing set based on a K-fold intersection method, and training the target detection model.
Specifically, the image position of the skier is identified based on the target detection image, the target person, that is, the image area of the skier is determined based on a predetermined frame selection ratio, the scene distribution ratio of the frame selection area is set, the frame selection area is determined based on the size of the image target person, and the skier position frame selection mark is made as the identification positioning image set.
Further, the identification positioning image set and the target detection image set are extracted, the target detection image set is used as level identification information, the identification positioning image set is used as level decision information, mapping correspondence is carried out on the identification positioning image set and the target detection image set, and therefore frame selection area identification and size positioning of the target person are carried out, and the sample data set is constructed. And equally dividing the sample data set into K groups, and determining K groups of sample data subsets. And further based on the K groups of sample data subsets, taking the K-1 groups of sample data subsets as training sets, taking the 1 groups of sample data subsets as test sets, training a neural network based on the training sets, acquiring the target detection model after training, further inputting the test sets into the target detection model, and obtaining a model output result and comparing deviation with the identification positioning image in the test sets to obtain the positioning deviation degree. And similarly, the sample quantity of the test set and the sample quantity of the training set are ensured to be unchanged, samples in the test set are adjusted, each group of sample data subsets can be used as the test set to carry out model training result test, deviation degree positioning is carried out respectively, a plurality of target detection models after training are determined, deviation degree comparison is further carried out, the minimum deviation degree is determined, and the corresponding training model is used as the finally determined target detection model, so that the operation mechanism of the model is effectively ensured to be better, and the detection accuracy is improved.
Step S300: configuring an image preprocessing space, and performing image preprocessing on the identification positioning image set based on the image preprocessing space to generate a re-identification image set;
further, as shown in fig. 2, the generating the re-identification image set, step S300 of the present application further includes:
step S310: identifying the skier position frame selection mark based on the identification positioning image, and obtaining a target identification result;
step S320: determining target pixel coordinates to cut the identification positioning image based on the target identification result, and obtaining a target demand image;
step S330: performing definition judgment on the target demand image to obtain an image meeting a definition threshold as a primary preprocessing image;
step S340: and carrying out image enhancement on the primary preprocessing image to serve as the re-identification image set.
Specifically, the image preprocessing space, namely the operation space for image adjustment, is configured, the image preprocessing process is determined, and the image preprocessing space is embedded. Specifically, the identification positioning image is extracted, skier position frame selection mark recognition is carried out in the image preprocessing space, so that image clipping area determination is carried out, and the image clipping area determination is used as the target recognition result. And (3) positioning target pixel coordinates based on the target recognition result, improving positioning accuracy, cutting the image on the basis, reserving a frame selection area, namely the area where the target person is located, and integrating the cut image to serve as the target demand image.
Further, a definition threshold is set, namely a critical limit value for image definition qualification judgment. And checking the target demand image and the definition threshold value, and providing an image with definition smaller than the definition threshold value, wherein the image has lower definition so as to cause identification deviation, and taking the image with definition greater than or equal to the definition threshold value as the primary preprocessing image. And further performing contour and contrast processing on the primary preprocessing image, performing color gamut adjustment, improving the visual effect of the image, improving the saliency of information, taking the enhanced primary preprocessing image as the re-identification image, and tamping the basis for the subsequent image feature identification extraction.
Step S400: generating a feature extraction model based on Object-reID network framework training based on the re-identified image set;
further, the generating a feature extraction model based on the re-recognition image set and the Object-reID network frame training, the step S400 of the present application further includes:
step S410: performing target area edge detection based on the re-identification image set, and determining an area set to be identified;
step S420: based on the region set to be identified, performing image convolution analysis to determine an image convolution feature set;
step S430: the Circle loss function is used as the loss function of the Object-Reid network framework;
step S440: and inputting an Object-Reid network framework according to the image convolution feature set and the region set to be identified, performing supervised training until a Circle loss function converges, and generating the feature extraction model.
Specifically, the re-identification image is extracted, edge detection is carried out on a target area, the target area is the coverage area of the re-identification image, edge detection is carried out to determine the frame selection edge of the image, so that the area positioning identification of the re-identification image is determined, identification errors are caused on the surface, and the detection result is used as the area set to be identified. And carrying out image convolution analysis on the region set to be identified, determining corresponding convolution characteristics through calculating convolution kernels, and further carrying out summarization and fusion on the convolution characteristics of the same target person to determine representative convolution characteristics as an even-numbered image convolution characteristic set. The Circle loss function is further used as the loss function of the Object-Reid network framework, and the contribution of the positive and negative samples to the gradients is controlled by adding a weight between the positive and negative samples, so that the model has more distinction degree in similarity discrimination, and the problem of low distinction degree of the skier image due to the same clothing is effectively improved.
Further, the image convolution feature set and the region set to be identified are input into an Object-Reid network framework for training, and supervised training is carried out until a Circle loss function converges, so that the feature extraction model is generated. And carrying out image feature recognition based on the feature extraction model, so that the output feature map can extract high-dimensional semantic information and rich spatial information. For the input skiing image, not only the abstract high-dimensional characteristics of the skiing image can be extracted, but also the spatial information characteristics for determining the body form and the action posture of the skiing image can be reserved. The similarity characteristics among different images of the same person and the differentiation characteristics among different persons are effectively improved, and the accuracy and the representativeness of characteristic extraction are improved.
Step S500: constructing an image feature processing module based on the target detection model, the image preprocessing space and the feature extraction model;
step S600: acquiring a retrieval request of a target user, and carrying out retrieval identification on a target image according to the image feature processing module to acquire target features;
specifically, a skiing image is collected, detection, processing and training are carried out, functional modeling is carried out to obtain the target detection model, the image preprocessing space and the feature extraction model, the image preprocessing space is placed in the target detection model and is placed in front of the feature extraction model, the image feature processing module is generated through link combination, and the image feature processing module is used as an auxiliary tool for image feature analysis.
Further, with the receiving of the search request of the target user, the search request is image search application information, the target user also has search additional information for the search request of a search executor, and the search additional information is an example detection image, namely, an image example of a search skier exists. Inputting the retrieval request into the image feature processing module, detecting and positioning skiers based on the target detection model, transmitting a positioning result into the image preprocessing space, performing image cutting, screening and enhancing processing, further transmitting the processed image to the feature extraction model, performing image convolution feature recognition and extraction based on a model training mechanism, taking a feature extraction result as the target feature, wherein the target feature is reference feature information to be subjected to image retrieval, taking the target feature as an index direction, performing associated image detection and identification, and the target feature is retrieval feature with representative and distinguishing index matched with a retrieval character and provides basis for subsequent image retrieval.
Step S700: traversing a storage database, and determining an image retrieval result based on the target feature, wherein the image retrieval result is expressed as a retrieval image-storage position-storage path, and the storage position and the storage path are additional output information.
Further, as shown in fig. 3, step S700 of the present application further includes:
step S710: traversing a storage database, and carrying out similarity matching based on the target features to obtain a similar sequence;
step S720: configuring a primary similarity threshold and retrieving an expected value;
step S730: determining an image sequence meeting the primary similarity threshold value as an image sequence to be evaluated based on the similarity sequence;
step S740: judging whether the magnitude of the image sequence to be evaluated is larger than the retrieval expected value or not;
step S750: when the primary similarity threshold value is larger than the primary similarity threshold value, the primary similarity threshold value is adjusted upwards, and threshold value judgment is carried out to finish updating of the image sequence to be evaluated;
step S760: and outputting an image retrieval result until the magnitude of the image sequence to be evaluated is smaller than or equal to the retrieval expected value.
In particular, the storage database is obtained, which stores rich ski images, and there are various storage locations and storage paths for the images. Traversing the storage database, taking the target features as matching identification information, performing similarity matching on each image in the storage database and the target features, determining a plurality of image similarities, further sequencing the plurality of image similarities from large to small, and synchronously sequencing corresponding images to generate the similarity sequence. The primary similarity threshold value and the retrieval expected value are further configured, the primary similarity threshold value is a similarity threshold value for image screening, and the retrieval expected value is a threshold value for limiting the number of screened images. And intercepting the similarity sequence based on the primary similarity threshold value, performing image reverse matching based on the intercepting result, and determining an image sequence meeting the similarity threshold value as the image sequence to be evaluated.
Further, determining the number of images of the image sequence to be evaluated, obtaining a sequence magnitude, and judging whether the sequence magnitude is larger than the retrieval expected value. When the similarity sequence is smaller than or equal to the retrieval expected value, the similarity sequence is indicated to meet the retrieval requirement, and the similarity sequence is used as the image sequence to be evaluated; when the search expected value is larger than the search expected value, the number of the search images meeting the primary similarity threshold is larger, and the images with higher similarity are determined as much as possible on the basis of ensuring that the number of the search images meets the standard, so that the search accuracy is further improved. Specifically, the primary similarity threshold is adjusted up, for example, the primary similarity threshold is adjusted up by one percentage point, interception and magnitude judgment of the similarity sequence are further performed, when the primary similarity threshold is still larger than the retrieval expected value, the primary similarity threshold is adjusted up again, adjustment of the primary similarity threshold is stopped until the magnitude of the finally determined image sequence is smaller than or equal to the retrieval expected value, the image sequence to be evaluated is updated, the finally determined image sequence to be evaluated is used as the image retrieval result, meanwhile, the storage position and the storage path of each image in the image retrieval result are determined and used as additional output information, the image retrieval result is marked, and a retrieval image-storage position-storage path sequence is generated and output. The matching degree of the image retrieval results can be effectively guaranteed within the retrieval requirements, and the possibility of deviation of retrieval abnormality is reduced.
Example 2
Based on the same inventive concept as a skier image retrieval method based on deep learning in the foregoing embodiments, as shown in fig. 4, the present application provides a skier image retrieval system based on deep learning, the system comprising:
the image set construction module 11 is used for connecting the Internet of things, taking a multi-dimensional scene as an index direction to acquire a skiing image set, and comprises an image storage position and a storage path to construct a target detection image set;
a target detection model training module 12, wherein the target detection model training module 12 is used for extracting an identification positioning image set based on the target detection image set and training a target detection model, and the identification positioning image set is provided with a skier position frame selection mark;
the image preprocessing module 13 is used for configuring an image preprocessing space, and performing image preprocessing on the identification positioning image set based on the image preprocessing space to generate a re-identification image set;
a feature extraction model training module 14, the feature extraction model training module 14 being configured to generate a feature extraction model based on Object-reID network framework training based on the re-recognition image set;
a construction module 15, wherein the construction module 15 is configured to construct an image feature processing module based on the target detection model, the image preprocessing space and the feature extraction model;
the feature acquisition module 16 is configured to acquire a search request of a target user, and perform search identification on a target image according to the image feature processing module to acquire a target feature;
the result determining module 17 is configured to traverse a storage database, and determine an image search result based on the target feature, where the image search result is expressed as a search image-storage location-storage path, and the storage location and the storage path are additional output information.
Further, the system further comprises:
the image acquisition module is used for connecting with the Internet and acquiring a skiing image set in a multi-dimensional scene as a target image set;
the image adjusting module is used for carrying out multistage light color adjustment and receptive field adjustment on the target image set and determining an amplified image set;
and the image set generation module is used for integrating the augmentation image set and the target image set to generate the target detection image set.
Further, the system further comprises:
the edge detection module is used for detecting the edge of the target area based on the re-identification image set and determining an area set to be identified;
the characteristic determining module is used for carrying out image convolution analysis based on the region set to be identified and determining an image convolution characteristic set;
the function embedding module is used for taking the Circle loss function as the loss function of the Object-Reid network framework;
and the model generation module is used for inputting an Object-Reid network frame according to the image convolution feature set and the region set to be identified, performing supervised training until the Circle loss function converges, and generating the feature extraction model.
Further, the system further comprises:
the image set extraction module is used for extracting the identification positioning image set based on the target detection image set;
the sample construction module is used for constructing a sample data set by taking the target detection image set as level identification information and the identification positioning image set as level decision information, wherein the target detection image set corresponds to the identification positioning image set one by one;
the sample dividing module is used for dividing the sample data set into K groups and obtaining K groups of sample data subsets;
and the model training module is used for dividing the K groups of sample data subsets into a training set and a testing set based on a K-fold intersection method and training the target detection model.
Further, the system further comprises:
the mark recognition module is used for recognizing the skier position frame selection mark based on the mark positioning image to obtain a target recognition result;
the image clipping module is used for determining target pixel coordinates to clip the identification positioning image based on the target identification result, and obtaining a target demand image;
the image screening module is used for judging the definition of the target demand image, and acquiring an image meeting a definition threshold as a primary preprocessing image;
and the image enhancement module is used for carrying out image enhancement on the primary preprocessing image and taking the primary preprocessing image as the re-identification image set.
Further, the system further comprises:
the sequence acquisition module is used for traversing the storage database, performing similarity matching based on the target features and acquiring a similar sequence;
the parameter configuration module is used for configuring a primary similarity threshold value and retrieving an expected value;
the sequence screening module is used for determining an image sequence meeting the primary similarity threshold value based on the similarity sequence and taking the image sequence as an image sequence to be evaluated;
the magnitude judgment module is used for judging whether the magnitude of the image sequence to be evaluated is larger than the retrieval expected value or not;
the sequence updating module is used for adjusting the primary similarity threshold upwards and judging the threshold to finish updating the image sequence to be evaluated when the magnitude of the image sequence to be evaluated is larger than the retrieval expected value;
and the result output module is used for outputting an image retrieval result until the magnitude of the image sequence to be evaluated is smaller than or equal to the retrieval expected value.
From the foregoing detailed description of a skier image search method based on deep learning, those skilled in the art can clearly understand that a skier image search method and system based on deep learning in this embodiment, for the device disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple, and relevant points refer to the method section.
The previous description of the disclosed embodiments is provided to enable any person skilled in the art to make or use the present application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the generic principles defined herein may be applied to other embodiments without departing from the spirit or scope of the application. Thus, the present application is not intended to be limited to the embodiments shown herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.
Claims (5)
1. A skier image retrieval method based on deep learning, the method comprising:
connecting the Internet of things, taking a multidimensional scene as an index direction to acquire a skiing image set, wherein the skiing image set comprises an image storage position and a storage path, and constructing a target detection image set; in particular, the method comprises the steps of,
connecting with the Internet, and collecting a skiing image set in a multi-dimensional scene as a target image set;
performing multistage light color adjustment and receptive field adjustment on the target image set to determine an augmented image set;
integrating and regulating the augmentation image set and the target image set to generate the target detection image set;
extracting an identification positioning image set based on the target detection image set, and training a target detection model, wherein the identification positioning image set is provided with a skier position frame selection mark;
configuring an image preprocessing space, and performing image preprocessing on the identification positioning image set based on the image preprocessing space to generate a re-identification image set;
generating a feature extraction model based on Object-reID network framework training based on the re-identified image set; the method for generating a feature extraction model based on Object-reID network framework training based on the re-identified image set comprises the following steps:
performing target area edge detection based on the re-identification image set, and determining an area set to be identified;
based on the region set to be identified, performing image convolution analysis to determine an image convolution feature set;
the Circle loss function is used as the loss function of the Object-Reid network framework;
inputting an Object-Reid network frame according to the image convolution feature set and the region set to be identified, performing supervised training until a Circle loss function converges, and generating the feature extraction model;
constructing an image feature processing module based on the target detection model, the image preprocessing space and the feature extraction model;
acquiring a retrieval request of a target user, and carrying out retrieval identification on a target image according to the image feature processing module to acquire target features;
traversing a storage database, and determining an image retrieval result based on the target feature, wherein the image retrieval result is expressed as a retrieval image-storage position-storage path, and the storage position and the storage path are additional output information.
2. The method of claim 1, wherein the extracting the set of identification localization images based on the set of target detection images trains a target detection model, the method comprising:
extracting the identification positioning image set based on the target detection image set;
taking the target detection image set as level identification information, and taking the identification positioning image set as level decision information to construct a sample data set, wherein the target detection image set corresponds to the identification positioning image set one by one;
dividing the sample data set into K groups, and obtaining K groups of sample data subsets;
and dividing the K groups of sample data subsets into a training set and a testing set based on a K-fold intersection method, and training the target detection model.
3. The method of claim 1, wherein the generating the set of re-identified images, the method comprising:
identifying the skier position frame selection mark based on the identification positioning image, and obtaining a target identification result;
determining target pixel coordinates to cut the identification positioning image based on the target identification result, and obtaining a target demand image;
performing definition judgment on the target demand image to obtain an image meeting a definition threshold as a primary preprocessing image;
and carrying out image enhancement on the primary preprocessing image to serve as the re-identification image set.
4. The method of claim 1, wherein the method comprises:
traversing a storage database, and carrying out similarity matching based on the target features to obtain a similar sequence;
configuring a primary similarity threshold and retrieving an expected value;
determining an image sequence meeting the primary similarity threshold value as an image sequence to be evaluated based on the similarity sequence;
judging whether the magnitude of the image sequence to be evaluated is larger than the retrieval expected value or not;
when the primary similarity threshold value is larger than the primary similarity threshold value, the primary similarity threshold value is adjusted upwards, and threshold value judgment is carried out to finish updating of the image sequence to be evaluated;
and outputting an image retrieval result until the magnitude of the image sequence to be evaluated is smaller than or equal to the retrieval expected value.
5. A deep learning based skier image retrieval system, the system comprising:
the image set construction module is used for connecting the Internet of things, taking the multidimensional scene as an index direction to acquire a skiing image set, and constructing a target detection image set by comprising an image storage position and a storage path; specifically, the image set construction module is used for connecting with the Internet, and collecting a skiing image set in a multi-dimensional scene as a target image set; performing multistage light color adjustment and receptive field adjustment on the target image set to determine an augmented image set; integrating and regulating the augmentation image set and the target image set to generate the target detection image set;
the target detection model training module is used for extracting an identification positioning image set based on the target detection image set and training a target detection model, wherein the identification positioning image set is provided with a skier position frame selection mark;
the image preprocessing module is used for configuring an image preprocessing space, and performing image preprocessing on the identification positioning image set based on the image preprocessing space to generate a re-identification image set;
the feature extraction model training module is used for generating a feature extraction model based on the re-identification image set and Object-reID network frame training; specifically, the feature extraction model training module is used for detecting the edge of the target area based on the re-identification image set and determining the area set to be identified; based on the region set to be identified, performing image convolution analysis to determine an image convolution feature set; the Circle loss function is used as the loss function of the Object-Reid network framework; inputting an Object-Reid network frame according to the image convolution feature set and the region set to be identified, performing supervised training until a Circle loss function converges, and generating the feature extraction model;
the building module is used for building an image feature processing module based on the target detection model, the image preprocessing space and the feature extraction model;
the feature acquisition module is used for acquiring a retrieval request of a target user, and carrying out retrieval identification on a target image according to the image feature processing module to acquire target features;
and the result determining module is used for traversing a storage database and determining an image retrieval result based on the target feature, wherein the image retrieval result is expressed as a retrieval image-storage position-storage path, and the storage position and the storage path are additional output information.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310120051.7A CN115858846B (en) | 2023-02-16 | 2023-02-16 | Skier image retrieval method and system based on deep learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310120051.7A CN115858846B (en) | 2023-02-16 | 2023-02-16 | Skier image retrieval method and system based on deep learning |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN115858846A CN115858846A (en) | 2023-03-28 |
| CN115858846B true CN115858846B (en) | 2023-04-21 |
Family
ID=85658123
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202310120051.7A Active CN115858846B (en) | 2023-02-16 | 2023-02-16 | Skier image retrieval method and system based on deep learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115858846B (en) |
Families Citing this family (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN118196196A (en) * | 2024-03-27 | 2024-06-14 | 北京大希科技有限公司 | Indoor image positioning method based on feature matching |
Citations (17)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105354548A (en) * | 2015-10-30 | 2016-02-24 | 武汉大学 | Surveillance video pedestrian re-recognition method based on ImageNet retrieval |
| CN106778464A (en) * | 2016-11-09 | 2017-05-31 | 深圳市深网视界科技有限公司 | A kind of pedestrian based on deep learning recognition methods and device again |
| CN108537136A (en) * | 2018-03-19 | 2018-09-14 | 复旦大学 | The pedestrian's recognition methods again generated based on posture normalized image |
| CN109214430A (en) * | 2018-08-15 | 2019-01-15 | 天津大学 | A kind of recognition methods again of the pedestrian based on feature space topology distribution |
| CN109255052A (en) * | 2018-08-29 | 2019-01-22 | 浙江工业大学 | A kind of three stage vehicle retrieval methods based on multiple features |
| CN109635634A (en) * | 2018-10-29 | 2019-04-16 | 西北大学 | A kind of pedestrian based on stochastic linear interpolation identifies data enhancement methods again |
| CN109886242A (en) * | 2019-03-01 | 2019-06-14 | 中国科学院重庆绿色智能技术研究院 | A kind of method and system that pedestrian identifies again |
| CN109902590A (en) * | 2019-01-30 | 2019-06-18 | 西安理工大学 | Pedestrian's recognition methods again of depth multiple view characteristic distance study |
| CN109948628A (en) * | 2019-03-15 | 2019-06-28 | 中山大学 | A kind of object detection method excavated based on identification region |
| CN110110601A (en) * | 2019-04-04 | 2019-08-09 | 深圳久凌软件技术有限公司 | Video pedestrian weight recognizer and device based on multi-space attention model |
| CN110321965A (en) * | 2019-07-10 | 2019-10-11 | 腾讯科技（深圳）有限公司 | The method and device that the training method of object weight identification model, object identify again |
| CN110942012A (en) * | 2019-11-22 | 2020-03-31 | 上海眼控科技股份有限公司 | Image feature extraction method, pedestrian re-identification method, device and computer equipment |
| CN111126249A (en) * | 2019-12-20 | 2020-05-08 | 深圳久凌软件技术有限公司 | Pedestrian re-identification method and device combining big data and Bayes |
| CN111460914A (en) * | 2020-03-13 | 2020-07-28 | 华南理工大学 | Pedestrian re-identification method based on global and local fine-grained features |
| CN112001251A (en) * | 2020-07-22 | 2020-11-27 | 山东大学 | Pedestrian re-identification method and system based on combination of human body analysis and clothing color |
| CN113326738A (en) * | 2021-05-06 | 2021-08-31 | 南京信息工程大学 | Pedestrian target detection and re-identification method based on deep network and dictionary learning |
| CN114547365A (en) * | 2022-02-22 | 2022-05-27 | 青岛海信网络科技股份有限公司 | Image retrieval method and device |
Family Cites Families (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| KR101879207B1 (en) * | 2016-11-22 | 2018-07-17 | 주식회사 루닛 | Method and Apparatus for Recognizing Objects in a Weakly Supervised Learning Manner |
| EP3851797A1 (en) * | 2020-01-14 | 2021-07-21 | Tata Consultancy Services Limited | Systems and methods for performing inclusive indoor navigation |
-
2023
- 2023-02-16 CN CN202310120051.7A patent/CN115858846B/en active Active
Patent Citations (17)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105354548A (en) * | 2015-10-30 | 2016-02-24 | 武汉大学 | Surveillance video pedestrian re-recognition method based on ImageNet retrieval |
| CN106778464A (en) * | 2016-11-09 | 2017-05-31 | 深圳市深网视界科技有限公司 | A kind of pedestrian based on deep learning recognition methods and device again |
| CN108537136A (en) * | 2018-03-19 | 2018-09-14 | 复旦大学 | The pedestrian's recognition methods again generated based on posture normalized image |
| CN109214430A (en) * | 2018-08-15 | 2019-01-15 | 天津大学 | A kind of recognition methods again of the pedestrian based on feature space topology distribution |
| CN109255052A (en) * | 2018-08-29 | 2019-01-22 | 浙江工业大学 | A kind of three stage vehicle retrieval methods based on multiple features |
| CN109635634A (en) * | 2018-10-29 | 2019-04-16 | 西北大学 | A kind of pedestrian based on stochastic linear interpolation identifies data enhancement methods again |
| CN109902590A (en) * | 2019-01-30 | 2019-06-18 | 西安理工大学 | Pedestrian's recognition methods again of depth multiple view characteristic distance study |
| CN109886242A (en) * | 2019-03-01 | 2019-06-14 | 中国科学院重庆绿色智能技术研究院 | A kind of method and system that pedestrian identifies again |
| CN109948628A (en) * | 2019-03-15 | 2019-06-28 | 中山大学 | A kind of object detection method excavated based on identification region |
| CN110110601A (en) * | 2019-04-04 | 2019-08-09 | 深圳久凌软件技术有限公司 | Video pedestrian weight recognizer and device based on multi-space attention model |
| CN110321965A (en) * | 2019-07-10 | 2019-10-11 | 腾讯科技（深圳）有限公司 | The method and device that the training method of object weight identification model, object identify again |
| CN110942012A (en) * | 2019-11-22 | 2020-03-31 | 上海眼控科技股份有限公司 | Image feature extraction method, pedestrian re-identification method, device and computer equipment |
| CN111126249A (en) * | 2019-12-20 | 2020-05-08 | 深圳久凌软件技术有限公司 | Pedestrian re-identification method and device combining big data and Bayes |
| CN111460914A (en) * | 2020-03-13 | 2020-07-28 | 华南理工大学 | Pedestrian re-identification method based on global and local fine-grained features |
| CN112001251A (en) * | 2020-07-22 | 2020-11-27 | 山东大学 | Pedestrian re-identification method and system based on combination of human body analysis and clothing color |
| CN113326738A (en) * | 2021-05-06 | 2021-08-31 | 南京信息工程大学 | Pedestrian target detection and re-identification method based on deep network and dictionary learning |
| CN114547365A (en) * | 2022-02-22 | 2022-05-27 | 青岛海信网络科技股份有限公司 | Image retrieval method and device |
Non-Patent Citations (1)
| Title |
|---|
| 郭迎春等.基于自适应融合网络的跨域行人重识别方法.自动化学报.2022,2744-2756. * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN115858846A (en) | 2023-03-28 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109857889B (en) | Image retrieval method, device and equipment and readable storage medium | |
| CN110414368B (en) | Unsupervised pedestrian re-identification method based on knowledge distillation | |
| CN111709311B (en) | Pedestrian re-identification method based on multi-scale convolution feature fusion | |
| CN109558810B (en) | Target person identification method based on part segmentation and fusion | |
| CN111582126B (en) | Pedestrian re-recognition method based on multi-scale pedestrian contour segmentation fusion | |
| CN107256017B (en) | Route planning method and system | |
| CN113989851B (en) | Cross-modal pedestrian re-identification method based on heterogeneous fusion graph convolution network | |
| CN111914642A (en) | Pedestrian re-identification method, device, equipment and medium | |
| CN111967527B (en) | Peony variety identification method and system based on artificial intelligence | |
| CN115858846B (en) | Skier image retrieval method and system based on deep learning | |
| CN110688955A (en) | Building construction target detection method based on YOLO neural network | |
| CN110443174A (en) | A kind of pedestrian's recognition methods again based on decoupling self-adaptive identification feature learning | |
| CN115620090A (en) | Model training method, low-illumination target re-recognition method and device and terminal equipment | |
| CN115909407A (en) | Cross-modal pedestrian re-identification method based on character attribute assistance | |
| CN117333948A (en) | End-to-end multi-target broiler behavior identification method integrating space-time attention mechanism | |
| CN118038494A (en) | Cross-modal pedestrian re-identification method for damage scene robustness | |
| CN112001280B (en) | Real-time and online optimized face recognition system and method | |
| CN116311357A (en) | Double-sided identification method for unbalanced bovine body data based on MBN-transducer model | |
| CN115050044B (en) | Cross-modal pedestrian re-identification method based on MLP-Mixer | |
| CN113780066B (en) | Pedestrian re-recognition method and device, electronic equipment and readable storage medium | |
| CN112488116B (en) | Scene understanding semantic generation method based on multi-mode embedding | |
| CN114627493A (en) | Gait feature-based identity recognition method and system | |
| CN111339904B (en) | Animal sperm image identification method and device | |
| CN114429605B (en) | Dynamic random target identification, positioning and tracking system in wide area environment | |
| CN117593717B (en) | Lane tracking method and system based on deep learning |
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 | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |