Weak and small target rapid high-precision identification system and method

Abstract

The invention relates to a fast and high-precision recognition system and method for a weak and small target, and aims to solve the problem that feature extraction and model construction based on data driving are difficult to directly apply a deep learning technology due to lack of enough effective sample data in weak and small target recognition under extreme conditions. A depth migration method based on cross-domain generative adversarial learning and compression excitation is provided to overcome the problem of insufficient data or low quality of weak and small target samples; due to the characteristics of weak detection targets and complex and diverse background environments, a target detection and identification method based on multi-instance mining and strong and weak supervision collaborative learning is provided to deal with the situation that the complex background of the weak small target is difficult to mark in details in an extreme environment; and a rapid and high-precision identification system for the weak and small target is constructed, and the rapid and high-precision identification of the weak and small target is realized by cooperatively utilizing the related technologies.

Description

Technical Field

[0001] The present invention relates to the field of target recognition, and in particular to a system and method for rapid and high-precision recognition of small and weak targets under extreme conditions. Background Art

[0002] Small target recognition under extreme conditions is the detection, tracking and identification of small targets through various sensors and algorithms under conditions of severe climate, complex terrain, low illumination, etc. It is of great significance in the fields of national defense and military, search and rescue, disaster monitoring, etc.

[0003] The current solution mainly uses related technologies based on generative adversarial networks (GANs), which generate new images with new shapes and features while retaining the main content of the original images by inputting a set of images into the GAN network to expand the data set. The deep learning model combined with the deep migration algorithm can make full use of the existing large-scale data sets to complete the training tasks of a small number of data models, relax the restrictions of the independent and identically distributed assumptions of training and test samples, and ensure the accuracy of the model.

[0004] The idea of ​​generative adversarial networks still requires a large amount of data to train generative adversarial networks. In target recognition tasks in extremely complex environments such as air, space, and sea, there are characteristics such as little effective data, high processing difficulty, and complex image annotation; conventional transfer learning has the disadvantages of imbalanced sample information between domains affecting cross-domain learning accuracy, and a large number of invalid features in high-dimensional feature channels. The above problems greatly limit the accuracy and efficiency of cross-domain transfer learning, and increase the difficulty and workload of manually screening effective data.

[0005] Due to the lack of sufficient effective sample data, it is difficult to directly apply deep learning technology to perform data-driven feature extraction and model construction for the recognition of weak and small targets under extreme conditions. At the same time, due to the weak and small characteristics of the targets and the complex and diverse background environment, the target detection algorithms that work well in everyday images often do not perform well in the detection imaging results under extreme environments. Summary of the invention

[0006] In response to the various challenges of weak target recognition in extreme environments mentioned above, the present invention proposes a deep migration method based on cross-domain generative adversarial learning and compressed excitation to overcome the problem of insufficient or low-quality weak target sample data; in addition, a target detection and recognition method based on multi-instance mining and strong and weak supervision collaborative learning is proposed to cope with the situation where it is difficult to fully label weak targets in complex backgrounds in extreme environments; finally, the above-mentioned related technologies are collaboratively used to construct a weak target fast and high-precision recognition system to achieve fast and high-precision recognition of weak targets.

[0007] 1. Deep migration method based on cross-domain generative adversarial learning and compressed excitation: Through cross-domain generative adversarial networks, feature pairs between image level and instance level are learned to deal with the imbalance problem of global domain migration and object representation migration. At the same time, a compressed excitation model is used to increase the weight of effective features, suppress invalid features, improve feature migration efficiency, and ensure recognition accuracy. The specific contents include:

[0008] (1) Constructing a feature extraction network for homogeneous and heterogeneous data migration: Considering the homogeneous and heterogeneous types of data, a domain separation network architecture is adopted to construct a homogeneous database within the source domain and a heterogeneous database between the target domain. Then, the encoder is used to learn the heterogeneous and homogeneous features between the source domain and the target domain to achieve feature conversion and migration between the source domain and the target domain data, thereby improving the generalization ability and adaptability of the model.

[0009] (2) High-dimensional feature transfer screening based on the compressed excitation model: The compressed excitation model is used to optimize feature channels and compensate for domain information. Through the lightweight SE module, the accuracy loss caused by domain shift is greatly reduced. At the same time, the interdependence between channels is explained, the channel response is adjusted, and a feature-based attention mechanism is realized.

[0010] (3) Object recognition based on cross-domain generative adversarial learning network: The source domain and target domain adaptation method is introduced, and a two-layer module is adopted. The image-level adaptive module of the segmentation pool is used to uniformly extract and align local block feature groups at different scales. Then, the instance-level adaptation module is used to semantically align the characteristics of the target group. Finally, the adaptation process of the two modules is regularized and stabilized by using the source model features, thereby ensuring the consistency of cross-domain adaptation and the accuracy of target positioning.

[0011] 2. Target detection and recognition method based on multi-instance mining and strong and weak supervision collaborative learning: Under the condition of image-level weak supervision annotation, multi-instance mining collaborative learning technology is used to achieve accurate mining of image instances and use them in the target detection and recognition method of strong and weak supervision collaborative learning, which ensures the target detection accuracy while reducing the requirements for data annotation accuracy. The specific contents are as follows:

[0012] (1) Object detection and recognition based on selective search and spatial pyramid pooling: The feature extraction network is pre-trained through a large number of image datasets, and the generalized texture features of the image are extracted. Then, the candidate boxes of the candidate regions are generated through the selective search method, and the feature vectors of the candidate regions are adaptively extracted through the spatial pyramid pooling method. The feature vectors of the candidate regions are processed by dual streams of classification and detection, and finally the cross entropy loss between image-level category prediction and category labeling is jointly optimized.

[0013] (2) Multi-instance mining based on spatial coupling and texture similarity: Instance objects of the same category usually have high spatial coupling and texture similarity. A spatial graph is used to describe the spatial relationship between candidate regions, and an appearance graph is used to describe the texture similarity between candidate regions. By combining the spatial coupling relationship and texture feature similarity, more instance objects of the same type are mined based on the prediction results of the weakly supervised target detector. In this way, in the network learning process based on the multi-instance learning method, target instances with less significant features can be detected and added to the training, thereby improving the expressiveness and robustness of the features.

[0014] (3) Model training algorithm based on strong and weak supervision collaborative learning: Both strong and weak supervisors are network models applied to target detection tasks. They should have similar semantic feature extraction structures, which provides a necessary foundation for strong and weak supervision collaborative learning. Therefore, the feature sharing method is used to combine strong and weak supervisors, which effectively enhances the consistency of the sample distribution representation of the two. At the same time, the prediction consistency loss is added to the strong supervision module. This enables the two to work independently or collaboratively, effectively improving the adaptability of the detection and recognition model in different scenarios and the upper limit of the algorithm accuracy.

[0015] 3. A system for fast and high-precision identification of small targets, including: resource management module, preprocessing and feature extraction module, pattern recognition model training module, detection and visualization module. The specific features of each module are as follows:

[0016] (1) Resource management module: completes data storage, query, access and other functions.

[0017] (2) Preprocessing and feature extraction module: completes data analysis and preprocessing, performs feature extraction and feature relationship extraction, and establishes the coupling relationship and feature mapping relationship between the image and typical targets.

[0018] (3) Pattern recognition model training module: The main function is to build a deep learning network layer and training optimizer, build a deep convolutional neural network structure with automatic target recognition function, and use the feature database to train the target recognition model.

[0019] (4) Detection and visualization module: It completes the tasks of target recognition, detection, and positioning through feature data and deep convolutional neural network calculations, and supports visualization analysis of detection data. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 This is a schematic diagram of a small target rapid and high-precision recognition system.

[0021] Figure 2 This is a schematic diagram of the deep migration method based on cross-domain generative adversarial learning and compressed excitation.

[0022] Figure 3 This is a schematic diagram of a target detection and recognition method based on multi-instance mining and strong and weak supervision collaborative learning. DETAILED DESCRIPTION

[0023] In order to better explain the technical solution, purpose and advantages of the present invention, the present invention will be described in detail in combination with the accompanying drawings and embodiments of the invention. Figure 1 Modules are divided in the figure, but in exceptional cases, the module division may be different from that in the figure, or the execution order may be different from that in the figure.

[0024] The present invention proposes a deep migration algorithm based on cross-domain generative adversarial learning and compressed excitation and a target detection and recognition method based on multi-instance mining and strong and weak supervision collaborative learning to solve the problems of few weak target samples and weak information.

[0025] 1. Deep migration method based on cross-domain generative adversarial learning and compressed excitation: First, build a deep model for target detection and recognition, design the compressed excitation module (SE) transfer learning architecture, adopt the segmentation pool method and candidate region sampling, and construct a feature pair mechanism; during the source domain model training process, build a domain adversarial network to achieve global domain migration of the image and feature isolation between target classes; use the SE module to dynamically realize high-dimensional feature weighting, complete effective feature screening and generation, and ultimately improve the model's cross-domain transfer learning ability and migration efficiency, ensuring target recognition accuracy and generalization ability.

[0026] (1) Constructing a feature extraction network for homogeneous and heterogeneous data migration: Using a domain separation network architecture, based on the source domain dataset X S With the target domain dataset X T , split the data into heterogeneous data With homogeneous data Heterogeneous data uses independent source domain encoders With the target domain encoder Homogeneous Data Sharing Encoder E c , construct the domain difference loss L difference and inter-domain consistency loss L similarity Design a deep transfer learning algorithm based on adversarial learning. First, initialize the feature extractor Θ and the classification regressor θ (classifier θ cls , regressor θ reg ), select sample set D, use gradient descent to optimize the model, where α is the learning rate, and get the Θ model after training, which is fixed for subsequent learning and testing.

[0027]

[0028]

[0029] In the parameter calculation of neurons, and Assuming that the i-th layer of the feature extractor Θ contains K neurons, it has K pairs of parameters, namely weights and biases. i，k , b i，k )}, then the parameters learned by the MTL model are K groups Assuming the input is X, the scale scaling and offset (SS) result is:

[0030]

[0031] (2) High-dimensional feature migration screening based on the compression excitation model: The lightweight SE module can greatly reduce the accuracy loss caused by domain shift, explain the interdependence between channels, adjust channel response, and realize feature-based attention mechanism. The further implementation steps are as follows:

[0032] S1: SE adapter consists of global pooling, fully connected, and RELU layers, where F avg is the global average pooling, F SE Combined with the FC+RELU+FC layer, an SE adapter group is constructed on this basis, and an SE adapter branch and inter-domain converter are added to each data domain so that it can select the adapter of the associated domain.

[0033] X SE =F SE (F avg (X)) (4)

[0034] S2: The SE adapter group performs feature screening in a priori known domains, and needs to actively set the domain conversion switch. To construct a universal SE adapter architecture, a universal representation space is formed by connecting the outputs of each domain adapter, where N is the number of adapters. For each adapter output, the adapter can be viewed as a nonlinear generalization of the filter bank for signal processing. Each nonlinear branch projects the output along a subspace that matches the statistics of a specific domain, and uses the attention mechanism to generate domain-sensitive weights to achieve multi-domain feature excitation.

[0035]

[0036] S3: The domain attention module is introduced to transfer feature projections using the general SE adapter structure, and the inter-domain information contribution is performed in a non-priori data-driven manner. In the domain attention module, the input feature map is first globally pooled, the spatial dimension is removed, and the output is constructed using linear plus softmax, where W DA is the softmax layer weight matrix, S DA Used to measure the output X of the universal SE adapterUSE , which in turn generates a domain activation response for channel multiplication F scale (·) Perform channel retuning.

[0037] S DA =E DA (X) = softmax(W DA F avg (X)) (6)

[0038] X DA =X USE S DA ∈R C×1 (7)

[0039]

[0040] (3) Object recognition based on cross-domain generative adversarial learning network: Introduce the source domain and target domain adaptation method, adopt a two-layer module, first use the image-level adaptive module of the segmentation pool to evenly extract and align local block feature groups at different scales, use the instance-level adaptation module to semantically align the characteristics of the target group, and stabilize the adaptation process of the two modules through source model feature regularization, thereby ensuring the consistency of cross-domain adaptation and target positioning accuracy. The further implementation steps are as follows:

[0041] S1: This case combines the classic target detection network model FasterR-CNN and the basic architecture of transfer learning, selects multi-scale anchor boxes and scaling factors, constructs multiple pairs of candidate region boxes, uses ROI pooling to generate fixed-size features, and performs segmentation pooling on large, medium and small scales to construct multiple local feature groups. l (f(x)),sp m (f(x)),Sp s (f(x))}, respectively represent the changes in the image-level domain such as the lighting environment, and then use the image-level adaptation model to perform multi-scale alignment and construct feature pairs in multi-scale to process image-level movement, where It consists of source domain samples. Composed of samples from the source domain and the target domain, the pairing method can effectively expand the limited target domain feature samples.

[0042]

[0043] g s ~Sp s (f(x T )), g t ~sp s (f(x T ) (10)

[0044] S2: In order to adapt the detection model, the goal of cross-domain adversarial learning is to align the two sets of constructed features, which is achieved by using the principle of generative adversarial learning. Through the mutual game between the feature generator and the domain discriminator, the domain difference between the two domains is minimized. Among them, the task of the feature generator is to realize the mutual conversion of source domain and target domain features to achieve feature alignment. At the same time, the domain discriminator attempts to distinguish whether the feature comes from the source domain or the target domain. The goal of the small-scale discriminator is to minimize a specific loss function, which focuses on accurately distinguishing between source-source feature pairs and source-target feature pairs. The goal of the overall image-level discriminator is to minimize The goal of the feature generator is to maximize

[0045]

[0046]

[0047]

[0048]

[0049] S3: To solve the domain drift problem of the target instance, the instance-level adaptation module is used to semantically align the target feature pairs, sample the candidate regions, construct the training data of the classifier and regressor, sample according to the positive and negative sample ratio, retain the candidate regions with higher IOU thresholds to ensure that they are closer to the real target, perform classification and regression, and obtain the source domain target feature set O. is and the target domain target feature set O it , pair them together for image block feature N i1 ={(n is , n is )}、N i2 ={(n is , n it )}, where n is ~O is , n it ~O it , the multi-way instance discriminator objective is to minimize and the feature generator objective is to minimize.

[0050]

[0051]

[0052] S4: In order to solve the problems of training instability and over-adaptation in adversarial learning, a strong regularization method is adopted. By forcing the adaptation model to produce consistent feature responses to the source model in the sense of l2 difference to the source input, the performance degradation caused by excessive updating of limited target samples is avoided. The source model feature regularization term is as follows: k is the number of foregrounds, the generated foreground area is penalized using the available presentation information, the source model is used to initialize the framework, and the entire framework is used for optimization.

[0053]

[0054]

[0055]

[0056] 2. Object detection and recognition method based on multi-instance mining and strong and weak supervision collaborative learning: Taking advantage of the high spatial coupling and texture similarity between instance objects of the same category, all instance targets in the image are accurately mined from image-level category annotations and integrated into the collaborative learning process based on the difference in target positioning between strong supervision and weak supervision. The specific contents are as follows:

[0057] (1) Weakly supervised object detection and recognition based on selective search and spatial pyramid pooling: First, the generalized texture features of the image are extracted by pre-training the feature extraction network with a large number of image datasets, and then the candidate region candidate boxes are generated by the selective search method. The feature vectors of the candidate regions are adaptively extracted by the spatial pyramid pooling method, and the feature vectors of the candidate regions are processed by dual streams of classification and detection. Finally, the cross entropy loss between image-level category prediction and category labeling is jointly optimized. The further implementation steps are as follows:

[0058] S1: In this case, a bottom-up search method is used. In order to search for as many candidate regions as possible, all small candidate regions can be searched through a small-scale window first, and then the small candidate regions can be screened out by comparing the color space distance and texture distance. According to the spatial position relationship, the relatively close similar candidate regions are synthesized into large candidate regions layer by layer.

[0059] S2: After the candidate region is generated, the spatial pyramid pooling method is used to divide the candidate region feature sub-map into sub-regions of different scales. After performing pooling operations on each sub-region, feature vectors at different scales are obtained. These feature vectors are concatenated and fused into the feature vector of the candidate region, and the classification score and positioning score are calculated. The two scores are combined to output the category prediction results of each candidate region.

[0060] S3: In this case, the total number of categories is C, and a binary category vector of length C is used to represent the image-level category information. If the category exists in the image, the corresponding binary bit is 1, otherwise the corresponding bit is 0. The cross entropy loss between the labeled category vector y makes these category prediction results close to the image-level category annotation results, and finally realizes the weakly supervised object detection and recognition method under the condition of only image-level category annotation.

[0061] (2) Multi-instance mining based on spatial coupling and texture similarity: Instance objects of the same category usually have high spatial coupling and texture similarity. A spatial graph is used to describe the spatial relationship between candidate regions, and an appearance graph is used to describe the texture similarity between candidate regions. Combining the spatial coupling relationship and texture feature similarity, more instance objects of the same type are mined based on the prediction results of the weakly supervised target detector, so that in the network learning process based on the multi-instance learning method, target instances with less significant features can be detected and added to the training, thereby improving the expressiveness and robustness of the features. Further implementation steps are as follows:

[0062] S1: Given an image candidate region P = {p1, p2, ..., p N} and image-level category annotation Y = y1, y2, ..., y C , the feature vector corresponding to the candidate region is F = {f1, f2, ..., f N}.

[0063] S2: For each category contained in the image, first obtain the most reliable candidate region output by weakly supervised target detection Then select the candidate regions with higher scores and compare them by intersection and union. Calculate the weight of any candidate region and , and thus construct a spatial graph of this class on the current image.

[0064] S3: The similarity measure d(f i , f j )=||f i -f j ||, construct the appearance map of the class on the current image. Thus, multiple candidate regions of the same category can be mined through the spatial map and appearance map.

[0065] S4: A weighted instance object cross entropy loss function is selected. By setting weights for candidate regions with different scores and integrating them into the cross entropy loss of each candidate region, the weights of instance objects with slightly lower scores can be increased, so that the weakly supervised object detection network can consider more obscure candidate regions, so that the network can learn more complete object instances at the same time, allowing the weakly supervised object detection method to obtain more accurate detection frames.

[0066] (3) The implementation steps of the model training algorithm based on strong and weak supervision collaborative learning are as follows:

[0067] S1: A weakly supervised target detection method based on multi-instance mining is selected as the weak supervisor, and a small sample target detection method based on domain adversarial generation and compression excitation is selected as the strong supervisor.

[0068] S2: The feature sharing method is combined with strong and weak supervisors to effectively enhance the consistency of sample distribution representation between the two. At the same time, the prediction consistency loss is added to the strong supervision module, where B w and B S They represent the number of targets detected by the strong supervisor and the weak supervisor respectively, C represents the number of target categories, and I ij represents the intersection-over-union weight, and β is a hyperparameter used to adjust the weights of each part. represents the cross entropy loss of strong and weak supervisor category prediction, Represents the information entropy of the target category detected by the strong and weak supervisors, Represents the target box detection loss of strong and weak supervisors.

[0069]

[0070] The solution provided by the present invention has the following effects: the proposed cross-domain generative adversarial network learns image-level and instance-level feature pairs to deal with the imbalance problem of global domain migration and object representation migration, and adopts a compressed excitation model to increase the weight of effective features, suppress invalid features, improve feature migration efficiency, and ensure recognition accuracy; multi-instance mining collaborative learning technology under image-level weak supervision annotation conditions can achieve accurate mining of image instances, and is used in a target detection and recognition method with strong and weak supervision collaboration to ensure target detection accuracy while reducing data security requirements.

[0071] The present invention also provides a system for fast and high-precision identification of small targets, including: a raw data resource management module; a data preprocessing and feature extraction module; a pattern recognition model training module; and a target identification and detection module. The specific composition and functions of each module are as follows:

[0072] The resource management module completes the functions of data storage, query, access, and recovery. It mainly includes the following sub-function modules: (1) Image data storage submodule: It uses tree structure and graph structure, adopts the latest big data storage and management technology, effectively organizes and stores target object information in different categories, and provides data export function. (2) Image data query submodule: supports multi-level classification and provides simultaneous display of image category and image content under the same category. It has information search function and provides the function of transferring image data between different categories. (3) Dataset concurrency control submodule: provides a concurrency control mechanism to allow multiple users to access image data in parallel. (4) Dataset fault recovery submodule: provides data set backup and recovery functions to improve the disaster recovery capability of the data set.

[0073] The preprocessing and feature extraction module completes data analysis and preprocessing, performs feature extraction and feature relationship extraction, and establishes the coupling relationship and feature mapping relationship between the image and the typical target. It mainly includes the following sub-function modules: (1) Data analysis submodule: Based on the CCD images, SAR images, optoelectronic videos and other data information acquired by the detector, the data is analyzed and preprocessed by using the method of expressing consistency of data in three-dimensional space, and the noise and invalid parts in the data are cleaned. (2) Classification and feature extraction submodule: It performs more accurate category identification and correction on the classified image data, extracts the feature vector of the image data, and reduces the redundancy of features through dimensionality reduction. (3) Feature relationship extraction submodule: It uses low-rank prior knowledge to establish the coupling relationship and feature mapping relationship between heterogeneous images and typical targets based on compressed sensing, nonlinear theory of measurement space, and continuous or discrete analysis on flow types.

[0074] The main function of the pattern recognition model training module is to build a deep learning network layer and training optimizer, build a deep convolutional neural network structure for automatic target recognition, and use the feature database to train the target recognition model. It mainly includes the following sub-function modules: (1) Network layer submodule: Construct the network layer module for deep learning, basic layer construction, including the construction of sub-functions such as convolutional layer, activation function layer, pooling layer, and fully connected layer. (2) Optimizer submodule: Builds the optimizer used during training and manages basic optimization algorithms. (3) Network architecture design submodule: Build a reasonable, timely, portable and stable convolutional deep neural network structure for automatic target recognition and positioning. (4) Back propagation training module: Use the target sample feature data that has been collected in the feature data set to train the target recognition model, and train according to the pre-defined model rules to obtain the desired recognition model.

[0075] The detection and visualization module mainly completes target recognition, detection and positioning through feature data and deep convolutional neural network calculations, and supports visual analysis of detection data. It mainly includes the following sub-function modules: (1) Target recognition submodule: After the target image is collected and the features are extracted, the features are brought into the recognition model for calculation to obtain the target recognition result. The data set can be imported synchronously while being visualized. (2) Target positioning submodule: The target positioning information is estimated through features and deep convolutional neural networks, and the recognition results are located on the map, making it convenient for operators to intuitively view the target's geographic location. (3) Information visualization submodule: supports the visualization analysis and management of detection data.

Claims

1. A system and method for rapid and high-precision identification of small targets, characterized in that: A deep migration method based on cross-domain generative adversarial learning and compressed excitation is used to overcome the problem of insufficient or low-quality sample data of weak targets; a target detection and recognition method based on multi-instance mining and strong and weak supervision collaborative learning is used to cope with the situation where weak targets and complex backgrounds in extreme environments are difficult to be fully labeled; a fast and high-precision recognition system for weak targets is constructed by collaboratively utilizing the above-mentioned related technologies to achieve fast and high-precision recognition of weak targets; the specific contents of each part are as follows: (1) A deep migration method based on cross-domain generative adversarial learning and compressed excitation, which uses a cross-domain generative adversarial network to learn feature pairs between the image level and the instance level to deal with the imbalance problem of global domain migration and object representation migration; at the same time, a compressed excitation model is used to increase the weight of effective features, suppress invalid features, improve feature migration efficiency, and ensure recognition accuracy; it is characterized by comprising the following steps: S1: Construct a feature extraction network for homogeneous and heterogeneous data migration, adopt a domain separation network architecture, and construct a homogeneous database within the source domain and the target domain, and a heterogeneous database between domains; then, the encoder is used to learn the heterogeneous and homogeneous features between the source domain and the target domain, thereby realizing feature conversion and migration between the source domain and the target domain data, so as to improve the generalization ability and adaptability of the model; S2: High-dimensional feature migration screening based on the compression excitation model, using the compression excitation model for feature channel optimization and domain information compensation, and through the lightweight SE module, greatly reducing the accuracy loss caused by domain shift, while explaining the interdependence between channels, adjusting channel response, and realizing feature-based attention mechanism; S3: Target recognition based on cross-domain generative adversarial learning network adopts a two-layer module, and uses the image-level adaptive module of the segmentation pool to uniformly extract and align local block feature groups at different scales; then, the instance-level adaptation module is used to semantically align the characteristics of the target group; finally, the adaptation process of the two modules is regularized and stabilized by using the source model features, thereby ensuring the consistency of cross-domain adaptation and the accuracy of target positioning; (2) A target detection and recognition method based on multi-instance mining and strong and weak supervision collaborative learning, which realizes accurate mining of image instances through multi-instance mining collaborative learning technology under image-level weak supervision annotation conditions, and is used in the target detection and recognition method of strong and weak supervision collaborative learning, while ensuring the target detection accuracy, reducing the requirements for data annotation accuracy; it is characterized by comprising the following steps: S1: Target detection and recognition based on selective search and spatial pyramid pooling. The feature extraction network is pre-trained through a large number of image datasets, and the generalized texture features of the image are extracted. Then, the candidate boxes of the candidate regions are generated through the selective search method, and the feature vectors of the candidate regions are adaptively extracted through the spatial pyramid pooling method. The feature vectors of the candidate regions are processed by dual streams of classification and detection, and finally the cross entropy loss between image-level category prediction and category labeling is jointly optimized. S2: Multiple instance mining based on spatial coupling and texture similarity, using a spatial graph to describe the spatial relationship between candidate regions, and using an appearance graph to describe the texture similarity between candidate regions; combining the spatial coupling relationship and texture feature similarity, mining more instance objects of the same type based on the prediction results of the weakly supervised target detector, so that in the network learning process based on the multiple instance learning method, target instances with less significant features can be detected and added to the training; S3: Model training based on strong and weak supervision collaborative learning, using feature sharing to combine strong and weak supervisors, effectively enhancing the consistency of sample distribution representation between the two; at the same time, adding prediction consistency loss in the strong and weak supervision modules; enabling the two to work independently or collaboratively, effectively improving the adaptability of the detection and recognition model in different scenarios and the upper limit of the algorithm accuracy; (3) A system for rapid and high-precision identification of small targets, characterized by comprising the following modules: a) Resource management module: complete data storage, query, access and other functions; b) Preprocessing and feature extraction module: complete data analysis and preprocessing, perform feature extraction and feature relationship extraction, and establish the coupling relationship and feature mapping relationship between the image and the typical target; c) Pattern recognition model training module: The main function is to build a deep learning network layer and training optimizer, build a deep convolutional neural network structure with automatic target recognition function, and use the feature database to train the target recognition model; d) Detection and visualization module: It completes the tasks of target recognition, detection and positioning through feature data and deep convolutional neural network calculations, and supports visual analysis of detection data.

Images