An infrared target detection method and device, electronic equipment and storage medium
By using an infrared target detection model based on an attention mechanism neural network, the problems of accuracy and efficiency in nighttime infrared target detection are solved, and efficient and accurate infrared target recognition is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- NANJING UNIV OF AERONAUTICS & ASTRONAUTICS
- Filing Date
- 2022-10-25
- Publication Date
- 2026-05-05
AI Technical Summary
Existing infrared target detection methods are ineffective at night or in low-light conditions, making it difficult to achieve efficient and accurate target recognition. In particular, deep learning-based methods are computationally complex and slow.
An infrared target detection model based on an attention mechanism neural network is adopted, including a feature extraction network, an attention module, a feature enhancement module, and a target classification and detection network. The detection accuracy is improved through feature fusion and attention mechanism.
It achieves efficient and accurate detection of infrared target images, improving the target recognition rate and detection accuracy at night.
Smart Images
Figure CN115496971B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of infrared target detection, and in particular to an infrared target detection method, apparatus, electronic device, and storage medium. Background Technology
[0002] In recent years, image acquisition and processing technologies based on infrared image sensors have developed rapidly in fields such as transportation, security, and environmental protection. Color visible light images (hereinafter referred to as color images) can achieve good target detection results under ideal lighting conditions, and deep learning techniques can be used to easily locate the specific position of targets of interest in color images and identify their categories. However, at night or in situations lacking sufficient light, target detection based on color images often fails to achieve the desired results, easily leading to missed detections, false detections, or even complete inoperability. Thermal infrared cameras are well-suited for imaging under such conditions because they can sense the thermal radiation emitted by targets and are not limited by lighting conditions. However, compared to color images, thermal infrared images have disadvantages such as low contrast and insufficient texture and edge features. These disadvantages make infrared target detection far more difficult than target detection in color images.
[0003] Currently, infrared target detection methods mainly include filtering-based methods, human visual system-based methods, low-rank sparse recovery-based methods, and deep learning-based methods. Among these four methods, filtering-based methods are suitable for single, uniform, continuous backgrounds and scenes with small target sizes; human visual system-based detection methods are mainly suitable for scenes where the target brightness is relatively high and there is a significant difference from the surrounding background; low-rank sparse recovery-based methods are suitable for complex and rapidly changing backgrounds, but have high computational complexity and are difficult to meet real-time requirements; deep learning-based methods mainly focus on two-stage algorithms, such as the R-CNN series. However, two-stage algorithms require first generating a proposal (a pre-selected box that may contain the object to be detected) and then performing infrared target category detection. These algorithms require multiple runs of the detection and classification process, which is complex and relatively slow. Summary of the Invention
[0004] The purpose of this invention is to provide an infrared target detection method, apparatus, electronic device, and storage medium that can perform low-cost, high-efficiency, and accurate detection of infrared target images.
[0005] To achieve the above objectives, the present invention provides the following solution:
[0006] In a first aspect, the present invention provides an infrared target detection method, comprising:
[0007] Acquire the currently acquired infrared target image;
[0008] The currently acquired infrared target image is input into the trained infrared target detection model for category detection;
[0009] The trained infrared target detection model is determined based on an attention mechanism neural network and a training dataset.
[0010] The training dataset includes multiple infrared target images and label data corresponding to each infrared target image; the label data represents the category of the infrared target image.
[0011] The attention mechanism neural network includes a feature extraction network, an attention module connected to the output of the feature extraction network, a feature enhancement module connected to the output of the attention module, and a target classification and detection network connected to the three outputs of the feature enhancement module.
[0012] The attention module is used to calculate the fusion coefficients of the original feature map output by the feature extraction network to determine the multi-scale feature map.
[0013] Optionally, the process of determining the trained infrared target detection model is as follows:
[0014] Build the training dataset;
[0015] Construct an attention mechanism neural network;
[0016] The attention mechanism neural network is trained based on the training dataset to obtain a trained infrared target detection model.
[0017] Optionally, the construction of the attention mechanism neural network specifically includes:
[0018] A feature extraction network is constructed; the feature extraction network is used to extract features from the infrared target image to obtain an original feature map;
[0019] Build an attention module;
[0020] A feature enhancement module is constructed; the feature enhancement module is used to perform feature fusion on the multi-scale feature map to obtain a fused feature map.
[0021] A target classification and detection network is constructed; the target classification and detection network is used to perform category detection based on the fused feature map and target box, determine the category of the infrared target image and obtain the prediction box.
[0022] Optionally, the feature extraction network is a Darknet network with a depth of 53; the construction process of the feature extraction network is as follows:
[0023] Set up the initial network;
[0024] The initial network was trained with weights using the ImageNet dataset to obtain the feature extraction network.
[0025] Optionally, the process of building the attention module is as follows:
[0026] Build the channel attention module;
[0027] The channel attention module is used for:
[0028] The original feature map is subjected to a non-dimensionality-reduction local cross-channel interaction strategy and an adaptive method for selecting the size of the one-dimensional convolution kernel. The original feature map is then subjected to global average pooling in the spatial dimension to obtain the average pooled feature vector.
[0029] After inputting the average pooling feature vector into the non-dimensionality-reduced GAP aggregated convolutional feature, the kernel size k is first adaptively determined, then one-dimensional convolution is performed, and then the channel attention is obtained by learning the Sigmoid function.
[0030] The channel attention is input to the feature enhancement module.
[0031] Optionally, the feature enhancement module specifically comprises:
[0032] Three feature layers are output from different positions of the attention module, located in the middle feature layer, the lower middle feature layer, and the bottom feature layer, respectively. The bottom feature layer is subjected to a 1×1 convolution to adjust the channels to obtain P5, which is then upsampled and combined with the lower middle feature layer. Then, CSPrayer is used to extract features to obtain P5_upsample. The feature layer of P5_upsample is subjected to a 1×1 convolution to adjust the channels to obtain P4, which is then upsampled and combined with the middle feature layer. Then, CSPrayer is used to extract features to obtain P4_upsample. The feature layer of P4_upsample is subjected to a 3×3 convolution to downsample, and after downsampling, it is stacked with P4. Then, CSPrayer is used to extract features to obtain P4_downsample. The feature layer of P4_downsample is subjected to a 3×3 convolution to downsample, and after downsampling, it is stacked with P5. Then, CSPrayer is used to extract features to obtain P5_downsample.
[0033] Optionally, the target classification and detection network specifically includes:
[0034] The fused feature map of the input is re-encoded using a 1×1 convolutional layer, a BN layer, and a SiLU activation function.
[0035] The recoded feature map is passed through two sets of two-layer 3×3 convolutions, BN layers, and SiLU activation functions to obtain the category and prediction box of the infrared target image. The first set of outputs is passed through a 1×1 convolution and sigmoid activation to determine the category of the infrared target contained in each feature point. The second set of outputs is passed through 1×1 convolution, sigmoid activation, and 1×1 convolution to determine the regression parameters of each feature point. After adjusting the regression parameters, the prediction box and whether each feature point contains an infrared target can be obtained.
[0036] Secondly, the present invention provides an infrared target detection device, comprising:
[0037] A data acquisition unit is used to acquire currently collected infrared target images;
[0038] A category detector is used to input the currently acquired infrared target image into a trained infrared target detection model for category detection;
[0039] The trained infrared target detection model is determined based on an attention mechanism neural network and a training dataset.
[0040] The training dataset includes multiple infrared target images and label data corresponding to each infrared target image; the label data represents the category of the infrared target image.
[0041] The attention mechanism neural network includes a feature extraction network, an attention module connected to the output of the feature extraction network, a feature enhancement module connected to the output of the attention module, and a target classification and detection network connected to the three outputs of the feature enhancement module.
[0042] The attention module is used to calculate the fusion coefficients of the original feature map output by the feature extraction network to determine the multi-scale feature map.
[0043] Thirdly, the present invention provides an electronic device including a memory and a processor, the memory being used to store a computer program, and the processor running the computer program to cause the electronic device to perform the infrared target detection method according to the first aspect.
[0044] Fourthly, the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the infrared target detection method as described in the first aspect.
[0045] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects:
[0046] This invention provides an infrared target detection method, apparatus, electronic device, and storage medium. The invention extracts features from infrared target images using a feature extraction network, and calculates fusion coefficients for the extracted features using an attention module to highlight infrared target features, thereby improving the accuracy of category recognition in infrared target images. Then, a feature enhancement module is used for feature fusion, which can reasonably and effectively fuse features of different scales, achieving good detection results for infrared targets of different sizes. Finally, the target classification and detection network is connected to the three outputs of the feature enhancement module, thereby achieving rapid and accurate detection of infrared target categories.
[0047] This invention uses an attention-based convolutional network to identify and detect infrared targets, improving the recognition rate of infrared targets at night and enhancing the accuracy of infrared target detection. Attached Figure Description
[0048] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0049] Figure 1 This is a flowchart of the infrared target detection method of the present invention;
[0050] Figure 2 This is a neural network framework diagram of the attention mechanism of the present invention;
[0051] Figure 3 This is a structural diagram of the channel attention module in the attention module of the present invention;
[0052] Figure 4 This is a structural diagram of the feature enhancement module of the present invention;
[0053] Figure 5 This is a diagram of the network structure for the target classification and detection of this invention. Detailed Implementation
[0054] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0055] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.
[0056] Among deep learning-based methods, two-stage algorithms, such as the R-CNN series, and one-stage algorithms, such as YOLO and SSD, differ primarily in that two-stage algorithms first generate a proposal (a pre-selected bounding box that may contain the object to be detected) and then perform infrared target category detection. These algorithms require multiple runs of the detection and classification process, making them complex and relatively slow. In contrast, one-stage algorithms directly extract features from the network to predict object classification and location, requiring only one input to the network to predict all bounding boxes, resulting in faster speed and higher accuracy. Therefore, this invention selects Yolox-m as a baseline and provides an efficient and accurate infrared target detection method, device, electronic device, and storage medium.
[0057] Example 1
[0058] like Figure 1 As shown, the infrared target detection method provided in this embodiment of the invention includes:
[0059] Step 101: Acquire the currently acquired infrared target image.
[0060] Step 102: Input the currently acquired infrared target image into the trained infrared target detection model for category detection.
[0061] The trained infrared target detection model is determined based on an attention mechanism neural network and a training dataset.
[0062] The training dataset includes multiple infrared target images and label data corresponding to each infrared target image; the label data is the category of the infrared target image.
[0063] The attention mechanism neural network includes a feature extraction network, an attention module connected to the output of the feature extraction network, a feature enhancement module connected to the output of the attention module, and a target classification and detection network connected to the three outputs of the feature enhancement module.
[0064] The attention module is used to calculate the fusion coefficients of the original feature map output by the feature extraction network to determine the multi-scale feature map.
[0065] exist Figure 1 In the aforementioned embodiment, the process of determining the trained infrared target detection model is as follows:
[0066] Step A: Construct a training dataset; the training dataset includes multiple infrared target images and label data corresponding to each infrared target image; the label data is the category of the infrared target image.
[0067] Step B: Construct an attention mechanism neural network.
[0068] Step C: Train the attention mechanism neural network based on the training dataset to obtain a trained infrared target detection model.
[0069] Furthermore, such as Figure 2 As shown, the construction of the attention mechanism neural network specifically includes:
[0070] Step B1: Construct a feature extraction network; the feature extraction network is used to extract features from the infrared target image to obtain an original feature map.
[0071] Step B2: Build the attention module.
[0072] Step B3: Construct a feature enhancement module; the feature enhancement module is used to perform feature fusion on the multi-scale feature map to obtain a fused feature map.
[0073] Step B4: Construct a target classification and detection network; the target classification and detection network is used to perform category detection based on the fused feature map and target box, determine the category of the infrared target image and obtain the prediction box.
[0074] exist Figure 1 In the described embodiment, the feature extraction network is a Darknet network with a depth of 53.
[0075] Furthermore, the construction of the feature extraction network specifically includes:
[0076] Step B11: Set up the initial network.
[0077] Step B12: Train the initial network using the ImageNet dataset to obtain the feature extraction network.
[0078] Furthermore, the construction of the attention module specifically involves:
[0079] Step B21: Build the channel attention module.
[0080] The channel attention module is used for:
[0081] The original feature map is subjected to a non-dimensionality-reduction local cross-channel interaction strategy and an adaptive method for selecting the size of the one-dimensional convolution kernel. The original feature map is then subjected to global average pooling in the spatial dimension to obtain the average pooled feature vector.
[0082] After inputting the average pooling feature vector into the non-dimensionality-reduced GAP aggregated convolutional feature, the kernel size k is first adaptively determined, then one-dimensional convolution is performed, and finally the channel attention is obtained by learning the Sigmoid function.
[0083] The channel attention is input to the feature enhancement module.
[0084] Furthermore, such as Figure 3 As shown, the channel attention module is specifically used for:
[0085] First, let's assume that one output from the Darknet network is... Where w, h, and c represent the width, height, and channel dimension, respectively. The feature map of the i-th channel is... The representation indicates that feature aggregation generates feature vectors Z = [z1, z2, ..., z] by employing global average pooling (GAP). s The nth element z of Z n The calculation formula is as follows:
[0086]
[0087] Then, information exchange between channels is achieved through one-dimensional convolution with a kernel size of k. The vector Z' = [z1', z2', ..., z] is obtained by one-dimensional convolution of Z. s '] and input it into the sigmoid function.
[0088] Finally, the attention weight vector z for each channel is... i 'and the feature map of the corresponding channel u i Multiplication yields a new characteristic expression u i ', u i '=z i 'u i Then the feature map is obtained. And input it into the next module.
[0089] Furthermore, the feature enhancement module is used to perform feature fusion on the multi-scale feature map using a feature enhancement method.
[0090] Among them, such as Figure 4 As shown, the feature enhancement module specifically consists of:
[0091] First, three feature layers are output from different positions of the attention module: the middle feature layer, the lower-middle feature layer, and the bottom feature layer. The bottom feature layer undergoes a 1×1 convolution to adjust the channels, resulting in P5. This P5 layer is then upsampled and combined with the lower-middle feature layer. CSPrayer is then used for feature extraction to obtain P5_upsample. The P5_upsample feature layer undergoes a 1×1 convolution to adjust the channels, resulting in P4. This P4 layer is then upsampled and combined with the middle feature layer. CSPrayer is then used for feature extraction to obtain P4_upsample. The P4_upsample feature layer undergoes a 3×3 convolution for downsampling. After downsampling, it is stacked with P4. CSPrayer is then used for feature extraction to obtain P4_downsample. The P4_downsample feature layer undergoes a 3×3 convolution for downsampling. After downsampling, it is stacked with P5. CSPrayer is then used for feature extraction to obtain P5_downsample.
[0092] Furthermore, such as Figure 5 As shown, the target classification and detection network specifically includes:
[0093] First, the input fused feature map is re-encoded using a 1×1 convolution, a Batch Normalization (BN) layer, and a SiLU activation function. The re-encoded feature map is then passed through two sets of two 3×3 convolutions, BN layers, and SiLU activation functions to obtain the category and predicted bounding box of the infrared target image. The first set of outputs is then passed through a 1×1 convolution and a sigmoid activation layer to determine the infrared target category contained in each feature point. The second set of outputs is then passed through a 1×1 convolution, a sigmoid activation layer, and a 1×1 convolution layer to determine the regression parameters for each feature point. After adjusting the regression parameters, the predicted bounding box and the determination of whether each feature point contains an infrared target are obtained. Finally, the three sets of output values are stacked.
[0094] Example 2
[0095] In order to execute the method corresponding to Embodiment 1 above and achieve the corresponding functions and technical effects, an infrared target detection device is provided below, specifically including:
[0096] The data acquisition unit is used to acquire the currently collected infrared target image.
[0097] A category detector is used to input the currently acquired infrared target image into a trained infrared target detection model for category detection.
[0098] The trained infrared target detection model is determined based on an attention mechanism neural network and a training dataset.
[0099] The training dataset includes multiple infrared target images and label data corresponding to each infrared target image; the label data is the category of the infrared target image.
[0100] The attention mechanism neural network includes a feature extraction network, an attention module connected to the output of the feature extraction network, a feature enhancement module connected to the output of the attention module, and a target classification and detection network connected to the three outputs of the feature enhancement module.
[0101] The attention module is used to calculate the fusion coefficients of the original feature map output by the feature extraction network to determine the multi-scale feature map.
[0102] Example 3
[0103] This invention provides an electronic device including a memory and a processor. The memory stores a computer program, and the processor runs the computer program to enable the electronic device to perform infrared target detection as described in Embodiment 1.
[0104] Alternatively, the aforementioned electronic device may be a server.
[0105] In addition, embodiments of the present invention also provide a computer-readable storage medium storing a computer program that, when executed by a processor, implements the infrared target detection of Embodiment 1.
[0106] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to the method section.
[0107] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.
Claims
1. An infrared target detection method, characterized in that, include: Acquire the currently acquired infrared target image; The currently acquired infrared target image is input into the trained infrared target detection model for category detection; The trained infrared target detection model is determined based on an attention mechanism neural network and a training dataset. The training dataset includes multiple infrared target images and label data corresponding to each infrared target image; the label data represents the category of the infrared target image. The attention mechanism neural network includes a feature extraction network, an attention module connected to the output of the feature extraction network, a feature enhancement module connected to the output of the attention module, and a target classification and detection network connected to the three outputs of the feature enhancement module. The attention module is used to calculate the fusion coefficients of the original feature map output by the feature extraction network to determine the multi-scale feature map; The feature enhancement module specifically comprises: outputting three feature layers from different positions of the attention module, located in the middle feature layer, the lower middle feature layer, and the bottom feature layer, respectively; the bottom feature layer undergoes a 1×1 convolution to adjust the channels to obtain P5, then is upsampled and combined with the lower middle feature layer, and then CSPrayer is used for feature extraction to obtain P5_upsample; the feature layer of P5_upsample undergoes a 1×1 convolution to adjust the channels to obtain P4, then is upsampled and combined with the middle feature layer, and then CSPrayer is used for feature extraction to obtain P4_upsample; the feature layer of P4_upsample undergoes a 3×3 convolution to downsample, and after downsampling, it is stacked with P4, and then CSPrayer is used for feature extraction to obtain P4_downsample; the feature layer of P4_downsample undergoes a 3×3 convolution to downsample, and after downsampling, it is stacked with P5, and then CSPrayer is used for feature extraction to obtain P5_downsample. The target classification and detection network specifically includes: recoding the input fused feature map through a 1×1 convolution, BN layer, and SiLU activation function; passing the recoded feature map through two sets of two 3×3 convolutions, BN layers, and SiLU activation functions to obtain the category and prediction box of the infrared target image; wherein, after the first set of outputs is passed through a 1×1 convolution and sigmoid activation, the infrared target category contained in each feature point is determined; after the second set of outputs is passed through a 1×1 convolution, sigmoid activation, and 1×1 convolution respectively, the regression parameters of each feature point are determined respectively, and after adjusting the regression parameters, the prediction box and whether each feature point contains an infrared target can be obtained.
2. The infrared target detection method according to claim 1, characterized in that, The process of determining the trained infrared target detection model is as follows: Build the training dataset; Construct an attention mechanism neural network; The attention mechanism neural network is trained based on the training dataset to obtain a trained infrared target detection model.
3. The infrared target detection method according to claim 2, characterized in that, The construction of the attention mechanism neural network specifically includes: A feature extraction network is constructed; the feature extraction network is used to extract features from the infrared target image to obtain an original feature map; Build an attention module; A feature enhancement module is constructed; the feature enhancement module is used to perform feature fusion on the multi-scale feature map to obtain a fused feature map. A target classification and detection network is constructed; the target classification and detection network is used to perform category detection based on the fused feature map and target box, determine the category of the infrared target image and obtain the prediction box.
4. An infrared target detection method according to claim 1 or 3, characterized in that, The feature extraction network is a Darknet network with a depth of 53; the construction process of the feature extraction network is as follows: Set up the initial network; The initial network was trained with weights using the ImageNet dataset to obtain the feature extraction network.
5. An infrared target detection method according to claim 1 or 3, characterized in that, The process of building the attention module is as follows: Build the channel attention module; The channel attention module is used for: The original feature map is subjected to a non-dimensionality-reduction local cross-channel interaction strategy and an adaptive method for selecting the size of the one-dimensional convolution kernel. The original feature map is then subjected to global average pooling in the spatial dimension to obtain the average pooled feature vector. After inputting the average pooling feature vector into the non-dimensionality-reduced GAP aggregated convolutional feature, the kernel size k is first adaptively determined, then one-dimensional convolution is performed, and then the channel attention is obtained by learning the Sigmoid function. The channel attention is input to the feature enhancement module.
6. An infrared target detection device, characterized in that, include: A data acquisition unit is used to acquire currently collected infrared target images; A category detector is used to input the currently acquired infrared target image into a trained infrared target detection model for category detection; The trained infrared target detection model is determined based on an attention mechanism neural network and a training dataset. The training dataset includes multiple infrared target images and label data corresponding to each infrared target image; the label data represents the category of the infrared target image. The attention mechanism neural network includes a feature extraction network, an attention module connected to the output of the feature extraction network, a feature enhancement module connected to the output of the attention module, and a target classification and detection network connected to the three outputs of the feature enhancement module. The attention module is used to calculate the fusion coefficients of the original feature map output by the feature extraction network to determine the multi-scale feature map; The feature enhancement module specifically comprises: outputting three feature layers from different positions of the attention module, located in the middle feature layer, the lower middle feature layer, and the bottom feature layer, respectively; the bottom feature layer undergoes a 1×1 convolution to adjust the channels to obtain P5, then is upsampled and combined with the lower middle feature layer, and then CSPrayer is used for feature extraction to obtain P5_upsample; the feature layer of P5_upsample undergoes a 1×1 convolution to adjust the channels to obtain P4, then is upsampled and combined with the middle feature layer, and then CSPrayer is used for feature extraction to obtain P4_upsample; the feature layer of P4_upsample undergoes a 3×3 convolution to downsample, and after downsampling, it is stacked with P4, and then CSPrayer is used for feature extraction to obtain P4_downsample; the feature layer of P4_downsample undergoes a 3×3 convolution to downsample, and after downsampling, it is stacked with P5, and then CSPrayer is used for feature extraction to obtain P5_downsample. The target classification and detection network specifically includes: recoding the input fused feature map through a 1×1 convolution, BN layer, and SiLU activation function; passing the recoded feature map through two sets of two 3×3 convolutions, BN layers, and SiLU activation functions to obtain the category and prediction box of the infrared target image; wherein, after the first set of outputs is passed through a 1×1 convolution and sigmoid activation, the infrared target category contained in each feature point is determined; after the second set of outputs is passed through a 1×1 convolution, sigmoid activation, and 1×1 convolution respectively, the regression parameters of each feature point are determined respectively, and after adjusting the regression parameters, the prediction box and whether each feature point contains an infrared target can be obtained.
7. An electronic device, characterized in that, The device includes a memory and a processor, the memory being used to store a computer program, and the processor running the computer program to cause the electronic device to perform the infrared target detection method according to any one of claims 1 to 5.
8. A computer-readable storage medium, characterized in that, It stores a computer program that, when executed by a processor, implements the infrared target detection method as described in any one of claims 1 to 5.
Citation Information
Patent Citations
Aircraft skin surface damage detection method and system based on deep learning
CN114565579A
Infrared target detection method based on feature fusion and attention mechanism
CN114694002A