An air target classification and identification method based on LSTM neural network
By combining LSTM neural networks with CNN visualization technology and Grad-CAM to generate feature maps, the problems of high computational cost and limited applicability in aerial target recognition are solved, and efficient classification and recognition of aerial targets in motion are achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SHANGHAI RADIO EQUIP RES INST
- Filing Date
- 2024-07-31
- Publication Date
- 2026-07-10
AI Technical Summary
Existing technologies for classifying and recognizing aerial targets involve large computational demands, have limited system portability, and struggle to effectively utilize image data from consecutive time points and the effects of pose for efficient recognition.
By employing an LSTM neural network-based approach, combined with CNN visualization technology and Grad-CAM feature map generation, and through key part annotation and feature classification of the target image, the LSTM neural network is used to simulate the local learning process, thereby enabling the identification of aerial targets in motion.
It achieves efficient classification and recognition of aircraft targets with very similar appearances, expands the scope of application, avoids the problem of inaccurate feature point selection, and takes into account the continuous time influence of moving aerial targets.
Smart Images
Figure CN118864980B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of electromagnetic scattering, and in particular relates to a method for classifying and recognizing aerial targets based on an LSTM neural network. Background Technology
[0002] During the flight of an aerial target, the radar emits electromagnetic wave signals that interact with the target and the background. After secondary radiation by the electromagnetic current induced by the target, a scattered echo is generated, producing a two-dimensional SAR image of the target. Aerial targets include a wide variety of aircraft targets. These aircraft targets are generally similar in structure, but differ in shape, type, and purpose. The differences between different models of the same type of aircraft may only exist in local parts. The method of classifying these slightly different aircraft targets is called fine-grained classification of aircraft targets. If these slightly different aircraft targets can be distinguished, it is of great significance in aircraft identification and other related fields.
[0003] However, current methods rely on massive computations, resulting in high computational costs. Furthermore, these network-based systems have limited portability and lack universal applicability. Moreover, for aerial targets considering their motion, how to comprehensively utilize continuous image data and the influence of the aerial target's attitude to achieve efficient identification of moving aerial targets remains to be studied. Summary of the Invention
[0004] The purpose of this invention is to provide an aerial target classification and recognition method based on LSTM neural network, which can achieve the effect of recognizing aerial targets in motion.
[0005] To achieve the above objectives, an aerial target classification and recognition method based on an LSTM neural network is provided for classifying and recognizing target images. The method includes: Step S1, constructing a visualization module for analyzing target images; Step S2, inputting the original target image into the visualization module for recognition, and performing feedback annotation on the original target image to obtain a target image with key parts annotated, wherein the target image with key parts annotated contains several annotated key parts; Step S3, inputting the target image with key parts annotated into an LSTM neural network model for feature classification and recognition to obtain feature classification and recognition results; Step S4, fusing and analyzing all feature classification and recognition results to obtain moving target recognition results.
[0006] Preferably, step S1 includes: constructing a visualization module for analyzing the target image using CNN visualization technology; the visualization module includes: a convolutional layer used to extract deep features from the target image; a global average pooling layer that receives data from the convolutional layer; the global average pooling layer optimizes the features extracted by the convolutional layer; a fully connected layer that receives data from the global average pooling layer; the fully connected layer performs dimensionality reduction on all the optimized data in the global average pooling layer; and a normalized exponential function layer that receives data from the fully connected layer; the normalized exponential function layer converts all the dimensionality-reduced data into a probability representation.
[0007] Preferably, step S2 includes: step S21, inputting the target image into the visualization module, wherein the target image generates n feature maps A in the convolutional layer. k Where k = 1, 2, ..., n: A k ∈R u×v ;R u×v Represent the feature map set; Step S22, for the feature map set R u×v Using a global average pooling layer, a fully connected layer, and a normalized exponential function layer, m categories are generated, where c = 1, 2, ..., m, and the probability value S for each category is obtained. c ;
[0008]
[0009] in, Let A be the weights of the feature map from the global average pooling layer to the normalized exponential function layer, where c is the index of the target class, and A is the weight of the target class. k It is the feature map output by the convolutional layer, f(A) k ) represents the dimensional features of the feature map output by the multiplication layer, and x represents the total number of feature maps of the same category; Step S23, the feature map A k And the probability value for each category is S c After probability filtering, the original target image is fed back to obtain a target image with key parts annotated. The target image with key parts annotated includes several annotated key parts.
[0010] Preferably, step S1 includes: constructing a visualization module for the target image using CNN visualization technology combined with Grad-CAM; the visualization module includes: a convolutional layer used to extract deep features from the target image; a global average pooling layer that receives data from the convolutional layer; the global average pooling layer optimizes the features extracted by the convolutional layer and performs dimensionality reduction on all optimized data; a normalized exponential function layer that receives data from the global average pooling layer; the normalized exponential function layer converts all dimensionality-reduced data into a probability representation.
[0011] Preferably, step S2 includes: step S21, inputting the target image into the visualization module, wherein the target image generates n feature maps A in the convolutional layer. k Where k = 1, 2, ..., n: A k ∈R u×v ;R u×v Represent the feature map set; Step S22, for the feature map set R u×v Using a global average pooling layer and a normalized exponential function layer, m categories are generated, where c = 1, 2, ..., m, and the probability value S for each category is obtained. c ;
[0012]
[0013] in, Let A be the weights of the feature map from the global average pooling layer to the normalized exponential function layer, where c is the index of the target class, and A is the weight of the target class. k It is the feature map output by the convolutional layer; f(A) k ) represents the dimensional features of the feature map output by the multiplication layer; x represents the total number of feature maps of the same category; step S23, convert feature map A k And the probability value for each category is S c After probability filtering, the original target image is fed back to obtain a target image with key parts annotated. The target image with key parts annotated includes several annotated key parts.
[0014] Preferably, step S3 includes: step S31, constructing an LSTM neural network model for analyzing the target image after key parts are annotated; step S32, inputting the target image after key parts are annotated into the LSTM neural network model and outputting the judgment result.
[0015] Preferably, the LSTM neural network model in step S31 includes, in sequence: an input layer, which receives input data and performs allocation operations on the data; an analysis layer, which receives data from the input layer and consists of several LSTM unit modules, in which several labeled key parts are screened, classified, and identified to obtain feature classification and identification results; and an output layer, which receives data from the analysis layer, integrates the preliminary data from the analysis layer, and outputs the feature classification and identification results.
[0016] Preferably, step S32 includes: step S321, inputting several labeled key parts into the input layer for allocation processing; step S322, the analysis layer receiving the labeled key parts from the input layer and performing forgetting processing according to a threshold, and judging the labeled key parts after forgetting processing; step S323, the output layer organizing the data and outputting it.
[0017] Preferably, step S322 includes the analysis layer receiving Z labeled key regions, z = 1, 2, ..., Z, and calculating the image data weight P of the z-th labeled key region. z :
[0018]
[0019] Where σ is the Sigmoid function; W is the weight matrix of the input layer; h z Theoretical output for marking key parts; b z This represents the bias of the vector representation of the key labeled parts. This involves creating vector representations of key labeled areas; and calculating the weight P for each key labeled area. z This process then determines whether to allow the image data of the marked key area to enter the judgment process.
[0020] Preferably, step S4 includes: using a Softmax classifier to perform fusion analysis on the feature classification and recognition results to obtain the final classification result, and then judging each moving target in the target image; specifically, if there is only one moving target, the judgment is made directly; when there are multiple moving targets, the feature classification and recognition results with a probability reaching the threshold need to be judged separately.
[0021] In summary, compared with the prior art, the aerial target classification and recognition method based on LSTM neural network provided by this invention has the following beneficial effects:
[0022] First, this invention analyzes Grad-CAM, which generates feature maps using CNN visualization technology, selects image regions of interest for target classification, and uses LSTM (Long Short-Term Memory) network to simulate the local learning process, thereby achieving excellent recognition and classification of target image data at continuous time points.
[0023] Secondly, the feature map generated by the CNN visualization technology after training avoids problems such as inaccurate selection of feature points of the target part, and takes into account the influence of continuous time of moving aerial targets, which can realize the classification and recognition of aerial targets and greatly expand the scope of application. Attached Figure Description
[0024] Figure 1 This is a flowchart of the present invention.
[0025] Figure 2 This is a schematic diagram illustrating the construction of the visualization module of the present invention.
[0026] Figure 3 (a) represents the computational logic in the order of global average pooling layer, fully connected layer, and normalized exponential function layer; Figure 3 (b) represents the operational logic in the order of the global average pooling layer and the normalized exponential function layer. Detailed Implementation
[0027] The following will be combined with the appendix in the embodiments of the present invention. Figure 1 ~Attached Figure 3 The technical solutions, structural features, objectives and effects achieved in the embodiments of the present invention will be described in detail.
[0028] It should be noted that the accompanying drawings are in a very simplified form and use non-precise proportions. They are only used to facilitate and clarify the purpose of illustrating the embodiments of the present invention, and are not intended to limit the implementation conditions of the present invention. Therefore, they have no substantial technical significance. Any modifications to the structure, changes in the proportional relationship, or adjustments to the size should still fall within the scope of the technical content disclosed in the present invention, provided that they do not affect the effects and objectives that the present invention can produce.
[0029] It should be noted that, in this invention, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only the expressly listed elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus.
[0030] like Figure 1 and Figure 2As shown, this invention provides an aerial target classification and recognition method based on LSTM (Long Short-term Memory) neural network, used for classifying and recognizing target images;
[0031] The target image is generally an image containing one or more moving targets, such as aircraft targets.
[0032] Regarding the acquisition of target images, radar can be used to acquire images of moving targets. For example, SAR (Synthetic Aperture Radar, an active Earth observation system that can be installed on aircraft, satellites, spacecraft, and other flight platforms) can be used to detect moving targets and obtain SAR images for subsequent processing. The choice of target image acquisition tools and methods can be made according to the actual situation.
[0033] After acquiring the target image, this invention acquires the key parts in the target image. The so-called key parts refer to the relevant elements in the target image used to identify the nature of the target (such as the component features of a moving target). The key parts of the target image can be obtained by refining the deep features in the target image, and then the key parts of the target image are judged to initially realize the identification of the target (general model and category).
[0034] Purification refers to removing interfering data, such as data generated by external cloud formations or light interference. This type of data has no computational significance and is therefore deleted.
[0035] However, after obtaining only the key parts of the target image, in order to ensure the accuracy of recognition, it is necessary to further clarify the key parts. At this time, the target image with the key parts already clarified is further processed (step S3) and fused analysis is performed (step S4) to obtain the recognition result (detailed model and category).
[0036] The following will describe in detail the aerial target classification and recognition method based on LSTM neural networks, which includes:
[0037] Step S1: Construct a visualization module for analyzing the target image;
[0038] The visualization module can be built using only CNN visualization technology (Convolutional Neural Networks), or in combination with other technologies, such as Grad-CAM (Gradient-weighted Class Activation Mapping); different combinations of these technologies will produce different visualization modules (see below for details).
[0039] Step S2: Input the original target image into the visualization module for recognition, perform feedback annotation on the original target image, and obtain the target image after key parts are annotated. The target image after key parts are annotated contains several annotated key parts.
[0040] Step S2 involves using the visualization module built in step S1 to analyze the target image, identify its key parts, and annotate them. Depending on the visualization module built in step S1, different calculation methods will be applied to the target image. Although different visualization modules / calculation methods exist, their purpose is to obtain a target image that includes several annotated key parts.
[0041] Step S3: Input the target image with key parts annotated into the LSTM neural network model for feature classification and recognition to obtain the feature classification and recognition results;
[0042] During the feature classification and recognition process, the identified data (several marked key parts) may be reduced and selected for recognition. The recognition result of a single marked key part will be output in the form of "key parts of the target image to be predicted - output result". Step S3 will also organize the data, and then step S4 will perform moving target recognition and judgment.
[0043] Step S4: Combine and analyze all feature classification and recognition results to obtain moving target recognition results;
[0044] The fusion analysis in step S4 refers to performing calculations on the processed data obtained in step S3 to obtain the recognition result of the moving target.
[0045] The following will describe steps S1 to S4 one by one. The visualization module built in step S1 is an important tool for generating key parts of the target image.
[0046] Here, we will first introduce the visualization module and its operating principle. When building the visualization module using CNN visualization technology, the visualization module includes the following components in sequence:
[0047] The convolutional layer is used to extract deep features from the target image. These deep features include relevant data on key parts of the target image, which require multiple refinements to obtain. The features are represented by each pixel in the target image through combination or independent methods, such as the texture and color of the target image. In this case, the deep features in the target image are represented in the convolutional layer as several matrices, each containing several pixels, with one matrix corresponding to one feature.
[0048] The global average pooling layer receives data from the convolutional layer. Its function is to optimize the features extracted from the convolutional layer. Because the deep features in the target image are represented by matrices, which are difficult to compute, the global pooling layer reduces the dimensionality of these matrices, shrinking them or even converting them into a single pixel block or point, thus reducing the computational complexity. In summary, the global average pooling layer effectively reduces the size of the matrices, thereby accelerating computation.
[0049] The fully connected layer receives data from the global average pooling layer. The function of the fully connected layer is to transform all the optimized matrices (or pixels) in the global average pooling layer into one-dimensional vectors (dimensionality reduction). At this time, the deep features in the target image are displayed in the form of corresponding one-dimensional vectors.
[0050] The normalized exponential function layer receives data from the fully connected layer; the normalized exponential function layer converts all one-dimensional vectors into a probability representation, and the normalized exponential function makes the probability value of each one-dimensional vector range between (0, 1), and the sum of the probabilities of all one-dimensional vectors is 1.
[0051] Therefore, the final result is that the deep features in the target image are represented in probability form, and the sum of all probability values is 1.
[0052] Therefore, in step S1, only CNN visualization technology is used to build the visualization module, which should sequentially include a convolutional layer, a global average pooling layer, a fully connected layer, and a normalized exponential function layer.
[0053] In a preferred embodiment, CNN visualization technology (Convolutional Neural Networks) combined with Grad-CAM (Gradient-weighted Class Activation Mapping) can also be used to build a visualization module for the target image.
[0054] Compared to the original CNN visualization technique, the construction of the visualization module will differ after incorporating Grad-CAM. The fully connected layers are removed, and the global average pooling layer assumes the functions of the fully connected layers (dimensionality reduction). The visualization module should then sequentially include a convolutional layer, a global average pooling layer, and a normalized exponential function layer, with the rest remaining unchanged. For example... Figure 3 The following are relevant examples, which will be explained in detail later.
[0055] Furthermore, when constructing visualization modules using CNN visualization technology (combined with Grad-CAM), computational training can be performed in advance, for example, by inputting known data into the visualization module to be trained. The known data refers to the results generated by a known target image and a pre-trained visualization module of the same type. Because the generated results are known, a usable visualization module can be obtained by continuously training and correcting the module. Specifically, the network training error (the difference between the result calculated by the normalized exponential function layer in the visualization module to be trained with a known target image and the result generated by a pre-trained visualization module of the same type with a known target image) can be calculated, and then backpropagated for continuous training and correction until excellent results are achieved.
[0056] Specifically, when using only CNN visualization technology to build the visualization module, the visualization module sequentially includes a convolutional layer, a global average pooling layer, a fully connected layer, and a normalized exponential function layer; its specific step S2 includes:
[0057] Step S21: Input the target image into the visualization module, where the target image generates n feature maps A in the convolutional layer. k Where k = 1, 2, ..., n: A k ∈R u×v ;R u×v Represents a set of feature maps;
[0058] The feature map A mentioned here k This represents the deep features in the aforementioned target image, indicating that there are n deep features in the target image;
[0059] Step S22, for the feature map set R u×v Using a global average pooling layer, a fully connected layer, and a normalized exponential function layer, m categories are generated, where c = 1, 2, ..., m, and the probability value S for each category is obtained. c ;
[0060]
[0061] in, Let A be the weights of the feature map from the global average pooling layer to the normalized exponential function layer, where c is the index of the target class, and A is the weight of the target class. k It is the feature map output by the convolutional layer, f(A) k ) represents the dimensional features of the feature map output by the layer, and x represents the total number of feature maps of the same category.
[0062] Step S23, transfer the feature map A k And the probability value for each category is S c After probability filtering, the original target image is fed back to obtain a target image with key parts annotated. The target image with key parts annotated includes several annotated key parts.
[0063] Specifically, after the normalized exponential function layer, the deep features in the target image are represented in probabilistic form. As mentioned earlier, because there are interfering objects, we select them according to the requirements and delete the deep features in the target image with a probability lower than a certain threshold (because their data volume is insufficient, indicating that they are interfering objects, and the threshold can be determined according to the actual situation). What remains are the key parts of the target image, which are then annotated to obtain the target image with several annotated key parts.
[0064] At this point, the several marked key parts obtained in step S23 are the areas of the target image that should be given special attention (key parts of the target image), which will be used as input to the LSTM neural network model in subsequent steps.
[0065] As mentioned earlier, the construction of the visualization module will differ after combining Grad-CAM. When using the CNN visualization technology Grad-CAM to construct the visualization module for the target image, the visualization module should sequentially include a convolutional layer, a global average pooling layer, and a normalized exponential function layer; this is equivalent to removing the fully connected layer. In this case, step S22 is equivalent to modifying the feature map set R. u×v m classifications are generated using a global average pooling layer and a normalized exponential function layer; for example... Figure 3 (a) and Figure 3 (b) As shown in the comparison. Figure 3 In (a), the left side represents the global average pooling layer, the middle rectangle represents the fully connected layer, and the right side represents the normalized exponential function layer; while Figure 3 In (b), the left side represents the global average pooling layer, the middle rectangle represents the fully connected layer, and the right side represents the normalized exponential function layer. Figure 3 In (a), the computation is performed sequentially through the global average pooling layer, the fully connected layer, and the normalized exponential function layer; Figure 3In (b), the operation skips the fully connected layer and goes directly from the global average pooling layer to the normalized exponential function layer. This operation logic is equivalent to using the global average pooling layer for dimensionality reduction (the global average pooling layer performs the function of the fully connected layer). The advantage of this approach is that it preserves the spatial and semantic information extracted by the convolutional layer, which is beneficial for retaining the information contained in the final output data.
[0066] So in Figure 3 In case (b), step S22 will be modified. In step S22, the feature map set R... u×v Using a global average pooling layer and a normalized exponential function layer, m categories are generated, where c = 1, 2, ..., m, and the probability value S for each category is obtained. c ;
[0067]
[0068] in, Let A be the weights of the feature map from the global average pooling layer to the normalized exponential function layer, where c is the index of the target class, and A is the weight of the target class. k It is the feature map output by the convolutional layer; f(A) k ) represents the dimensional features of the feature map output by the layer; x represents the total number of feature maps of the same category.
[0069] The biggest difference between the two calculation methods lies in the weights of the feature map from the global average pooling layer to the normalized exponential function layer. This data is different; the following steps are also the same, except for step S22.
[0070] Specifically, step S3 includes:
[0071] Step S31: Construct an LSTM neural network model for analyzing the target image after key parts are annotated; wherein the LSTM neural network model includes an input layer, an analysis layer and an output layer in sequence;
[0072] Step S32: Input the target image with key parts annotated into the LSTM neural network model and output the judgment result.
[0073] The construction of this LSTM neural network model requires training. The LSTM neural network model constructed after training has three layers. The first layer is the input layer, which is used to receive the target image after the key parts are annotated. Its function is to receive the input data (the target image after the key parts are annotated). In this layer, the image will be assigned to which LSTM unit module in the analysis layer.
[0074] The second layer is the analysis layer, which is the main layer for analysis. It receives data from the input layer (the target image after the key parts have been labeled). It contains several LSTM unit modules, which are also the executors of the specific operations. Each LSTM unit module is configured with different operation logic. The input layer assigns the aforementioned labeled key parts to the corresponding (one or more) LSTM unit modules according to certain rules (such as classification probability values), and performs filtering, classification and recognition on the labeled key parts to obtain the feature classification and recognition results.
[0075] The key to the LSTM neural network model lies in the selective processing of key parts by the LSTM unit modules. Specifically, the Sigmoid function (also known as the S-shaped growth curve, often used as the activation function of neural networks) partially forgets some of the received labeled key parts (by selecting a threshold to determine which key parts of the target image will not be included in the subsequent calculation). Then, the labeled key parts that need to be predicted are classified and identified to obtain preliminary data.
[0076] The third layer is the output layer, which integrates the preliminary data obtained from the analysis layer to output the feature classification and recognition results. Since the data analyzed by the analysis layer is scattered data in the form of [key parts of the target image to be predicted - output results], the output layer organizes and summarizes the data and associates it with the image to facilitate the processing and judgment in the subsequent step S4.
[0077] Therefore, step S32 includes:
[0078] Step S321: Input several key annotated parts into the input layer for allocation processing;
[0079] Step S322: The analysis layer receives the labeled key parts from the input layer, performs forgetting processing according to the threshold, and judges the labeled key parts after forgetting processing.
[0080] Step S323: The output layer organizes the data and outputs it.
[0081] In step S322, specifically, the Sigmoid function can be used to calculate the threshold and remove parts of the image that do not need to be judged.
[0082] Here is an example: suppose there are Z key labeled regions, z = 1, 2, ..., Z. The image data weight P of the z-th key labeled region is calculated. z :
[0083]
[0084] In the formula, σ is the Sigmoid function (an activation function used in neural networks, also known as the S-shaped growth curve); W is the weight matrix of the input layer; h z Theoretical output for marking key parts; b z This represents the bias of the vector representation of the key labeled parts. This is a vector representation of the key labeled parts; that is, it can be calculated by assigning a weight P to each key labeled part. z This process then determines whether to allow the image data of the marked key area to enter the judgment process.
[0085] After discarding some of the key labeled parts, the remaining key labeled parts are judged, and the result of "Key parts of the target image to be predicted - output result" is generated one by one.
[0086] Furthermore, when using an LSTM neural network model, training is inevitably required. Therefore, this section introduces a method for error assessment during LSTM neural network model training, which involves inputting known data into the LSTM neural network model to be trained. The known data refers to the target image and the results generated by a pre-trained LSTM neural network model of the same type. Because the generated results are known, a usable visualization module can be achieved through continuous training and calibration. Specifically, the off-target error between the output data of the LSTM neural network model to be trained (third layer) and the output data that should be generated by the known data is calculated. In addition, since the output data is presented in the form of [key parts of the target image to be predicted - output result], it needs to be processed to convert it into the form of the accuracy or overlap rate with the actual target, which is called the overlap probability value, to facilitate subsequent calculations.
[0087] Specifically as follows:
[0088]
[0089] Where y is the probability of overlap between the output data of a pre-trained LSTM neural network model of the same type and known data (generally, it should be between 0.9 and 1). L represents the probability of overlap between the output data and the known data in the LSTM neural network model to be trained, where L is the number of overlapping output probability values. By continuously using off-target values... This is used to assess the performance of the LSTM neural network model and continuously optimize it until the off-target effect is achieved. To achieve a value required for practical use.
[0090] Step S4 mainly fuses the feature classification and recognition results in step S3. Since the output layer has already organized and summarized the data and mapped it to the image, the recognition result can be obtained directly by mapping the image here.
[0091] For example, a Softmax classifier (a classifier suitable for multi-class problems, which converts the output of a neural network into a probability distribution, providing a probability value for each class) can be used for classification to obtain the final classification result; then, the various moving targets in the target image can be identified. When using a Softmax classifier, if there is only one moving target, there is only one type of target, which is easy to identify; when there are multiple moving targets, multiple similar probability distributions will appear, and the feature classification results with probabilities reaching a certain threshold need to be judged separately.
[0092] Although the present invention has been described in detail through the preferred embodiments above, it should be understood that the above description should not be considered as a limitation of the present invention. Various modifications and substitutions to the present invention will be apparent to those skilled in the art after reading the above description. Therefore, the scope of protection of the present invention should be defined by the appended claims.
Claims
1. A method for classifying and recognizing aerial targets based on an LSTM neural network, characterized in that, The method for classifying and recognizing target images includes: Step S1: Construct a visualization module for analyzing the target image; Step S2: Input the original target image into the visualization module for recognition, perform feedback annotation on the original target image, and obtain the target image after key parts annotation. The target image after key parts annotation includes several annotated key parts. Step S3: Input the target image with key parts annotated into the LSTM neural network model for feature classification and recognition to obtain the feature classification and recognition results; Step S4: Combine and analyze all feature classification and recognition results to obtain moving target recognition results; Step S3 includes: Step S31: Construct an LSTM neural network model for analyzing the target image after key parts are annotated; Step S32: Input the target image with key parts annotated into the LSTM neural network model and output the judgment result; The LSTM neural network model in step S31 includes, in sequence: The input layer receives input data and performs allocation operations on the data. The analysis layer receives data from the input layer and consists of several LSTM unit modules. The LSTM unit modules filter, classify, and identify several labeled key parts to obtain feature classification and identification results. The output layer receives data from the analysis layer, integrates the preliminary data from the analysis layer, and outputs the feature classification and recognition results. Step S32 includes: Step S321: Input several key annotated parts into the input layer for allocation processing; Step S322: The analysis layer receives the labeled key parts from the input layer and performs forgetting processing according to the threshold, and then judges the labeled key parts after forgetting processing. Step S323: The output layer organizes the data and outputs it. Step S322 includes: The analysis layer receives Z labeled key regions, z=1, 2...Z, and calculates the image data weight P of the z-th labeled key region. z : in, For the Sigmoid function; This is the weight matrix of the input layer; Theoretical output for marking key parts; This represents the bias of the vector representation of the key labeled parts. This involves creating vector representations of key labeled areas; and calculating the weight P for each key labeled area. z This process then determines whether to allow the image data of the marked key area to enter the judgment process.
2. The aerial target classification and recognition method based on LSTM neural network according to claim 1, characterized in that, Step S1 includes: constructing a visualization module for analyzing the target image using CNN visualization technology; The visualization module includes: Convolutional layer, which is used to extract deep features from the target image; The global average pooling layer receives data from the convolutional layer; its function is to optimize the features extracted from the convolutional layer. The fully connected layer receives data from the global average pooling layer; the function of the fully connected layer is to perform dimensionality reduction processing on all the optimized data in the global average pooling layer. The normalized exponential function layer receives data from the fully connected layer; the normalized exponential function layer transforms all the dimensionality-reduced data into a probability representation.
3. The aerial target classification and recognition method based on LSTM neural network according to claim 2, characterized in that, Step S2 includes: Step S21: Input the target image into the visualization module, where the target image generates n feature maps in the convolutional layer. Where k = 1, 2, ..., n: ; Represents a set of feature maps; Step S22, for the feature map set The system generates m categories, c=1, 2, ..., m, using a global average pooling layer, a fully connected layer, and a normalized exponential function layer, and obtains the probability value for each category. ; in, Here, c represents the weights of the feature map from the global average pooling layer to the normalized exponential function layer, and c is the index of the target class. It is the feature map output by the convolutional layer. This represents the dimensional features of the feature map output by the multiplicative layer, where x represents the total number of feature maps of the same category. Step S23, generate the feature map and the probability value of each category After probability filtering, the original target image is fed back to obtain a target image with key parts annotated. The target image with key parts annotated includes several annotated key parts.
4. The aerial target classification and recognition method based on LSTM neural network according to claim 1, characterized in that, Step S1 includes: using CNN visualization technology combined with Grad-CAM to build a visualization module for the target image; The visualization module includes: Convolutional layer, which is used to extract deep features from the target image; The global average pooling layer receives data from the convolutional layer; its function is to optimize the features extracted from the convolutional layer and reduce the dimensionality of all optimized data. The normalized exponential function layer receives data from the global average pooling layer; the normalized exponential function layer transforms all the dimensionality-reduced data into a probability representation.
5. The aerial target classification and recognition method based on LSTM neural network according to claim 4, characterized in that, Step S2 includes: Step S21: Input the target image into the visualization module, where the target image generates n feature maps in the convolutional layer. Where k = 1, 2, ..., n: ; Represents a set of feature maps; Step S22, for the feature map set Using a global average pooling layer and a normalized exponential function layer, m categories are generated, where c = 1, 2, ..., m, and the probability value for each category is obtained. ; in, Here, c represents the weights of the feature map from the global average pooling layer to the normalized exponential function layer, and c is the index of the target class. It is the feature map output by the convolutional layer; This represents the dimensional features of the feature map output from the multiplicative layer; x represents the total number of feature maps of the same category. Step S23, generate the feature map and the probability value of each category After probability filtering, the original target image is fed back to obtain a target image with key parts annotated. The target image with key parts annotated includes several annotated key parts.
6. The aerial target classification and recognition method based on LSTM neural network according to claim 1, characterized in that, Step S4 includes: The Softmax classifier is used to fuse and analyze the feature classification and recognition results to obtain the final classification result, and then to determine each moving target in the target image; Specifically, if there is only one moving target, a judgment is made directly; if there are multiple moving targets, the feature classification and recognition results with probabilities reaching the threshold need to be judged separately.