A camouflaged target detection and recognition method based on deep neural network

Abstract

The present invention proposes an end-to-end camouflaged target detection and recognition method based on a deep neural network. The method designs a dual-path neural network of 'segmentation-recognition'. The 'segmentation' path uses the overall image information of the target, takes the anti-attention convolution module as the main body, and adds the Receptive Field Block to ensure that the network obtains a larger receptive field. Its overall design adopts the U-net network structure, which can more effectively and accurately capture the color, texture and other deep visual features of the camouflaged target. The 'recognition' path adopts a dual-branch Resnet structure to convert the target pixel information into semantic information to identify the specific target type. Experimental results show that this method can effectively solve the problems of conventional target detection and recognition methods that are difficult to detect and completely segment camouflaged targets. In a variety of complex environments such as oceans, jungles, snow, deserts and other areas, it can effectively separate and identify a variety of camouflaged targets such as animals, plants, human bodies, military facilities, etc.

Description

Technical Field

[0001] The present invention relates to the field of target detection and computer vision technology, and in particular to a camouflaged target detection and recognition method based on a deep neural network. Background Art

[0002] Camouflage refers to the way or means that things use various methods to hide the truth and show falsehood in order to achieve a certain purpose. Camouflage technology is widely present in nature and human society. For example, various animals and plants use camouflage to capture food or avoid natural enemies; camouflage technology is widely used in modern military as a very important anti-reconnaissance means to provide false information and confuse the opponent. According to the characteristics of camouflage, camouflage types can be divided into: similar color and texture camouflage, false target camouflage, smoke camouflage, cover camouflage, etc. The current mainstream camouflage target detection and processing methods can be divided into camouflage target detection methods based on the combination of digital image processing and machine learning and camouflage target detection methods based on deep learning.

[0003] The camouflage target detection method combining digital image processing and machine learning is mainly studied around the camouflage method of similar color and texture camouflage, which has relatively obvious characteristics. The main feature of the method is that the digital image processing method is used to artificially extract some camouflage information features, such as texture and color discontinuity, brightness difference, characteristic corners, etc. Then, machine learning theories such as PCA (Principal Component Analysis) dimensionality reduction and SVM (support vector machines) are used for feature classification, and the classification detection results are obtained after image synthesis processing. For example, Hongxing Guo et al. proposed a camouflage foreground and background division method based on Bayesian classification and Gaussian mixture model in visual monitoring applications. By averaging the sequence of video frames to reduce the variance of the camouflage target image frame, the accuracy of camouflage object detection in the video is improved. For example, Galun et al. proposed a new texture segmentation method for camouflage targets. This method first identifies the sub-element types of textures, characterizes them using relevant parameters, and applies them to texture camouflage target detection, achieving a high recognition accuracy.

[0004] Compared with traditional methods, camouflage target detection based on deep learning is applicable to a wider range of camouflage types and has a relatively high recognition accuracy. In the deep learning-based method, the digital image processing in the traditional method is replaced by a deep neural network to extract camouflage information features. Compared with the traditional method, this method has stronger adaptability and is more conducive to discovering deeper camouflage features. In addition, the feature classifier of the deep learning method is also easier to perform nonlinear activation than the traditional method, so it can better handle nonlinear classification and regression problems, and it is manifested in better image segmentation effect and higher recognition accuracy in the camouflage target detection task. For example, Jia-Xing Zhao et al. proposed EGNet, which improves the unfavorable expression of salient edge information and salient object information in the full convolutional neural network, and uses salient edge features to help salient object features more accurately locate the edge of the object. Its related methods and results can also be applied to camouflage target detection. For example, Deng-Ping Fan et al. established the COD10K camouflage target data set and proposed the SINet network, which incorporates the ideas and methods of search and recognition, has strong versatility, and has achieved good results in the accuracy of camouflage target recognition and segmentation accuracy.

[0005] Due to the complexity of the image background and target information of the disguised target, the existing neural network structure still cannot meet the high-precision segmentation requirements, and there will be problems such as incomplete segmentation or the segmented area not matching the actual area. In addition, most of the existing disguised target detection models focus on the segmentation of disguised target images, and there are relatively few models that perform disguised target category recognition. Summary of the invention

[0006] Purpose of the invention: The technical problem to be solved by the present invention is to provide a camouflaged target detection and recognition method based on deep neural network in view of the shortcomings of the prior art. The method can accurately detect whether a camouflaged target exists in a variety of complex background environments only through image information, and can also more accurately segment the binary area of ​​the camouflaged target from the image, and identify a variety of camouflaged objects such as people, animals, military targets, etc.

[0007] In order to solve the above technical problems, the present invention discloses a camouflaged target detection and recognition method based on a deep neural network, comprising the following steps:

[0008] Step 1: Constructing a disguised target data set: including collecting disguised target data, image preprocessing, manual annotation and data set division; the data set division refers to dividing the data set into training set data, verification set data and test set data;

[0009] Step 2: Building a disguised target detection model: The disguised target detection model is used to input a disguised target image and output a disguised target image segmentation and recognition result;

[0010] Step 3: Train the disguised target detection model: Use the training set data to train the constructed disguised target detection model. During the training process, periodically use the validation set data to conduct preliminary tests on the training results, and adjust the details of the disguised target detection model based on the feedback results of the preliminary tests.

[0011] Step 4: Test the disguised target detection model: Use the test set data to test the detection effect of the disguised target detection model obtained after training in step 3, and calculate the test evaluation index.

[0012] Furthermore, step 1 includes:

[0013] Step 1.1, collecting camouflaged target data under complex backgrounds, where the collected data type can be static image data or dynamic video data; the complex backgrounds include oceans, jungles, snowy areas, and desert areas;

[0014] Step 1.2, data preprocessing: the collected camouflaged target data is uniformly cropped into a single-frame color image of size 480×480×3, and then normalized, that is, the image RGB pixel values ​​are normalized to the mean and variance of the ImageNet public dataset to obtain the camouflaged target image;

[0015] Step 1.3, manual labeling: manually label the actual area of ​​the camouflaged target image, add a label, and mark the target type;

[0016] Step 1.4, data set division: divide the data set processed in step 1.3 into training set data, validation set data and test set data, and the proportion of the three sets in the total data set are 70%, 15% and 15% respectively.

[0017] Through step 1, a complete disguised target dataset is constructed to lay the data foundation for subsequent training and testing of the disguised target detection model.

[0018] Furthermore, the disguised target detection model in step 2 includes a segmentation network and a recognition network:

[0019] The segmentation network is a fully convolutional neural network, which is used to extract the ROI (region of interest) area of ​​the camouflaged target image; the input of the segmentation network is the camouflaged target image, and the output is a binary black-and-white image containing the camouflaged target ROI area; the segmentation network includes a front half network, a middle half network and a back half network, the front half network encodes the camouflaged target image to obtain a feature map of the camouflaged target image; the obtained feature map is input into the middle half network for feature fusion to obtain a feature fusion result; the feature fusion result is decoded by using the back half network to obtain a binary black-and-white image containing the camouflaged target ROI area;

[0020] The recognition network is designed with a double-branch residual network Resnet, which is used to identify the type of camouflaged targets; the input of the recognition network is the camouflaged target image and the binary black-and-white image containing the camouflaged target ROI area obtained by the segmentation network, and the output is the camouflaged target recognition result.

[0021] The camouflaged target detection model designs different network structures for two different task types. Compared with the conventional border detection method, the segmentation and recognition method can provide a more accurate detection area. And the output of the segmentation network is input into the recognition network, which is conducive to the recognition network to quickly grasp the regional high-dimensional features and improve the accuracy of the recognition network.

[0022] Furthermore, the first half of the segmentation network adopts a U-net structure, wherein the U-net structure refers to a U-shaped transformation in which the resolution of the obtained feature map changes from large to small and then to large; the left half of the U-net structure is composed of a Resnet-50 backbone network obtained by Imagenet pre-training, which is used to extract and encode multi-scale feature information of the camouflaged target image to obtain four preliminary encoding features; the right half of the U-net structure is composed of four cascaded ordinary convolution blocks, which are used to refine the classification and decoding of the four preliminary encoding features obtained by the left half of the network to obtain four preliminary decoding features.

[0023] The U-net design used can significantly expand the network's expansion path and improve the model's ability to locate camouflaged target information. At the same time, the backbone network is connected to the corresponding size of the upsampling part to reduce the loss of information in the network flow.

[0024] Furthermore, the middle part of the segmentation network is a receptive field block (RFB), which splits each of the four outputs of the right half of the U-net structure into four branches, each of which uses convolution kernels of different sizes and convolution methods with different void rates; then, convolution and multiplication are used to fuse the features of the four branches to obtain four feature fusion results.

[0025] Furthermore, the second half of the segmentation network is an anti-attention module, which inputs the four preliminary decoding features obtained from the first half of the segmentation network and the four feature fusion results obtained from the middle part to obtain four binary black-and-white images containing the camouflaged target ROI area, including: the preliminary decoding features obtained from the first half of the segmentation network are used to obtain the attention coefficient through the Sigmoid activation function, and the attention coefficient is subtracted from the value 1 to obtain the anti-attention coefficient; the anti-attention coefficient is Hadamard-producted with the feature fusion result obtained from the middle part of the segmentation network, and the calculation result is input into the single-layer volume network to obtain the binary black-and-white image containing the camouflaged target ROI area.

[0026] The middle part and the latter part of the segmentation network are refinement modules designed at the end of the network, which are used to further improve the quality of the network output image. Conventional refinement modules mostly use an encoding-decoding structure, which will stack more convolutional layers and take up a lot of computation; and conventional methods mostly only use the final layer of the network for refinement and adjustment, and do not use the information of other feature maps, and the information used is too single. The middle part of the segmentation network improves the receptive field module, which can enhance the local nonlinear performance of the network. The latter part introduces an anti-attention mechanism, which enables the network to emphasize or select important information of the disguised processing object and suppress some irrelevant detail information.

[0027] Furthermore, the recognition network adopts a dual-branch Resnet structure, the camouflaged target image is directly input into one Resnet branch, and the binary black-and-white image containing the camouflaged target ROI area obtained by the segmentation network is input into another Resnet branch after passing through a convolutional network amplification channel; after merging the feature maps output by the two Resnet branches, feature fusion is performed through a continuous three-layer convolutional network, and then input into a two-layer BP (Back Propagation) neural network to obtain a camouflaged target recognition result.

[0028] The recognition network simultaneously utilizes the original image information and the binary graphic information output by the segmentation network, thereby avoiding learning features directly from complex camouflaged target background information; making the recognition network easier to train and converge.

[0029] Furthermore, when training the disguised target detection model in step 3, the loss function of the segmentation network is a hybrid loss function, which is expressed by the following formula:

[0030]

[0031] Where L n Represents the loss function of the segmentation network part, pred i represents the binary black-and-white image containing the disguised target ROI area output by the segmentation network, gt represents the true value image, and l bce_loss , l iou_loss and l ssim_loss They respectively represent the BCE (Binary Cross Entropy) loss, IOU (Intersection Over Union) loss and SSIM (Structural Similarity) loss between the binary black and white image and the camouflaged area image, λ1, λ2 and λ3 represent the loss weight coefficients, and the default values ​​are all 1. i=1, 2, 3, 4 correspond to the four output values ​​of the segmentation network.

[0032] The hybrid loss function can effectively evaluate the differences between images from multiple perspectives, including pixels, regions, and the whole, and effectively alleviate the negative impact of different object sizes on segmentation performance.

[0033] Furthermore, when training the camouflaged target detection model in step 3, the recognition network is trained using the cross entropy loss function L rn For training, the formula is as follows:

[0034] L m = l cross_entropy (predk,label)

[0035] Among them, predk represents the vector corresponding to the target type output by the recognition network, and label represents the vector corresponding to the actual target type; the training adopts the Adam optimizer, and sets the initial hyperparameters batch size batch_size = 32, learning rate learning_rate = 0.0001, and total training rounds epoch = 40.

[0036] Furthermore, the test evaluation indicators described in step 4, for the segmentation network, use the mean absolute error MAE (Mean Absolute Error), enhanced matching index E-MEASURE (Enhanced-alignment Measure) and F-MEASURE indicators to evaluate the segmentation accuracy of the disguised target; for the recognition network, the average recognition accuracy of the model is used to evaluate the accuracy of the disguised target recognition.

[0037] Beneficial effects:

[0038] 1. The embodiment of the present application uses a deep neural network to realize the recognition of camouflaged targets and the segmentation of camouflaged areas. Compared with the prior art that mostly uses borders to directly detect camouflaged targets, the method proposed in the present application can not only segment camouflaged areas but also identify the types of camouflaged targets, and its functions are relatively comprehensive.

[0039] 2. The neural network model proposed in the embodiment of the present application only uses image information as input, does not require the assistance of other sensor devices, and the network structure adopts an end-to-end design. Therefore, the embodiment of the present application has the characteristics of simple deployment and easy practical application.

[0040] 3. The embodiment of the present application designs a novel deep neural network model, introduces a receptive field module and an anti-attention mechanism, and improves the loss function. The above approach enables the model to achieve good results in both the accuracy of camouflaged target image segmentation and the accuracy of target recognition, and is highly competitive among existing camouflaged target detection methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] The present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments, and the above and / or other advantages of the present invention will become more clear.

[0042] Figure 1 It is an overall flow chart of the camouflaged target detection and identification method proposed in the embodiment of the present application;

[0043] Figure 2 This is an example of the information that needs to be annotated when building a camouflaged target dataset;

[0044] Figure 3 It is a structural diagram of the segmentation network part of the camouflaged target detection method proposed in the embodiment of the present application;

[0045] Figure 4 It is the structural diagram of the Receptive Field Block module;

[0046] Figure 5 This is the structural diagram of the Reverse Attention module;

[0047] Figure 6 It is a structural diagram of the identification network part of the disguised target detection method proposed in the embodiment of the present application;

[0048] Figure 7 It is a demonstration of the detection effect of the disguised target detection method proposed in the embodiment of the present application on part of the data in the COD10K public disguised target data set;

[0049] Figure 8 It is a schematic diagram of an application scenario of a camouflaged target detection and recognition method based on a deep neural network provided in an embodiment of the present application.

[0050] Fig. 9 It is a schematic diagram of qualitative comparison of the segmentation effect of the disguised target detection method proposed in the embodiment of the present application on a public disguised target data set and other target segmentation algorithms. DETAILED DESCRIPTION

[0051] The technical solution of the present invention is further described below in conjunction with the accompanying drawings and specific examples.

[0052] The present invention proposes an end-to-end camouflaged target detection and recognition method based on a deep neural network. The method adopts a dual-path design of "segmentation-recognition", which can accurately segment the binary image of the camouflaged target in a relatively complex environment and accurately identify the type of the camouflaged target. Test results show that this method can effectively separate and identify a variety of camouflaged targets such as animals, plants, people, military facilities, etc. in a variety of complex background environments such as oceans, jungles, snow, deserts, etc. Figure 8 The figure is a schematic diagram of an application scenario of a camouflaged target detection and recognition method based on a deep neural network provided in an embodiment of the present application.

[0053] like Figure 1 As shown, the present invention proposes a camouflaged target detection and recognition method based on a deep neural network, and the specific steps of establishing the method are as follows:

[0054] Step 1: Construct a disguised target dataset. Different datasets need to be established according to different disguised detection targets, so that more targeted training can be carried out to improve the accuracy of target detection. The method of constructing a dataset is shown in the following four sub-steps. You can also select public datasets of disguised targets such as COD10K and CAMO for model training. The models trained with these datasets are more versatile, but due to the large number of disguised targets, the relevant performance indicators of the models may be relatively general.

[0055] Step 1.1, collect camouflaged target data. Find a suitable camouflaged background and camouflaged object, and collect its RGB color photo or video information. The image needs to be large enough to ensure that its size can cover the entire camouflaged target. Convert all collected image videos into single-frame pictures, and ensure that there are at least 500 images of each type of camouflaged target that needs to be detected, so as to ensure that the deep neural network established in the embodiment of the present application has enough data to fit the features. Camouflaged backgrounds include oceans, jungles, snow and desert areas, and camouflaged objects include animals, plants, people, military facilities, etc.

[0056] Step 1.2, data preprocessing. Since the input of the neural network usually has a fixed form, the images collected in step 1.1 need to be processed uniformly. Here, the images are first cropped to a size of 480×480×3. Then, in order to eliminate the influence of color type and brightness intensity on the features, the cropped images are standardized, that is, the three pixel values ​​of RGB in the image are normalized to the mean [0.485, 0.456, 0.406] and variance [0.229, 0.224, 0.225] of Imagenet, and the camouflaged target image is obtained.

[0057] Step 1.3: Manual labeling. For each camouflaged target image obtained by preprocessing in step 1.2, manually label a binary image gt of the same size and a category label label. Figure 2 As shown in the figure, the black area in GT is the non-camouflage area, and the white area is the camouflage area. This step requires that the annotation should be fine enough to clearly distinguish the camouflage and non-camouflage areas to ensure the reliability of the overall information of the dataset. Photoshop or labelme is generally used to annotate the camouflage area.

[0058] Step 1.4, data set division. Randomly shuffle all the data sets annotated in step 1.3, and then take out 70% as training set data, 15% as validation set data, and 15% as test set data.

[0059] Step 2: Build a camouflaged target detection model. Use a deep learning framework to build a deep neural network model, which is mainly divided into two parts: segmentation network and recognition network.

[0060] In this embodiment, the segmentation network is a fully convolutional neural network, which is used to extract the ROI area of ​​the disguised target image. Figure 3 ; The input of the segmentation network is a disguised target image, and the output is a binary black-and-white image containing a disguised target ROI area; the segmentation network includes a front half network, a middle half network and a back half network, the front half network encodes the disguised target image to obtain a feature map of the disguised target image; the obtained feature map is input into the middle half network for feature fusion to obtain a feature fusion result; the feature fusion result is decoded by the back half network to obtain a binary black-and-white image containing a disguised target ROI area;

[0061] The recognition network adopts a dual-branch Resnet design to identify the types of camouflaged targets. The main structure is shown in Figure 6 The input of the recognition network is the disguised target image and the binary black-and-white image containing the disguised target ROI area obtained by the segmentation network, and the output is the disguised target recognition result.

[0062] The first half of the segmentation network adopts the U-net structure, which means that the resolution of the obtained feature map changes from large to small and then to large. Figure 3 As shown in the figure, the left half of the U-net structure is composed of the Resnet-50 backbone network obtained by Imagenet pre-training, which is used to extract and encode the multi-scale feature information of the image and obtain preliminary encoding features; the right half of the U-net structure is composed of four cascaded ordinary convolution blocks, which are used to refine the classification and decoding of the four preliminary encoding features obtained by the left half of the network and obtain four preliminary decoding features.

[0063] The middle part of the segmentation network is the receptive field block, such as Figure 4 As shown in the figure, each of the four outputs of the right half of the U-net structure is further forked into four branches, and each branch uses convolution kernels of different sizes and convolution methods with different void rates; then convolution and multiplication are used to fuse the features of the four branches to obtain four feature fusion results.

[0064] The second half of the segmentation network is the anti-attention module, which is implemented using a convolutional layer and is used to decode the feature maps obtained by the first half of the network, such as Figure 5 As shown. The four preliminary decoding features obtained by the first half of the segmentation network and the four feature fusion results obtained by the middle part are input to obtain four binary black-and-white images containing the camouflaged target ROI area, including: the preliminary decoding features obtained by the first half of the segmentation network are used to obtain the attention coefficient through the Sigmoid activation function, and the attention coefficient is subtracted from the value "1" to obtain the anti-attention coefficient; the anti-attention coefficient is used to perform the Hadamard product with the feature fusion result obtained by the middle part of the segmentation network, and the calculation result is input into the single-layer volume network to obtain the decoding result, that is, the binary black-and-white image containing the camouflaged target ROI area.

[0065] The recognition network adopts a dual-branch Resnet structure, such as Figure 6 As shown in the figure, there are two inputs. The camouflaged target image information is directly input into one Resnet branch, and the binary black-and-white image containing the camouflaged target ROI area obtained by the segmentation network is input into another Resnet branch after being amplified through the convolutional network channel; after the feature maps output by the two Resnet branches are merged, the features are fused through a continuous three-layer convolutional network, and then input into a two-layer BP neural network to obtain the camouflaged target recognition result.

[0066] Step 3: Train the disguised target detection model. Using the deep learning framework, read the data into the disguised target detection model to start training. The segmentation network part uses a mixed loss function, which is expressed as follows:

[0067]

[0068] Where L n Represents the loss function of the segmentation network part, pred i Represents the binary black-and-white image of the camouflaged target ROI area output by the segmentation network, gt is the camouflaged area image marked above, and l bce_loss , l iou_loss and l ssim_loss They represent the BCE loss, IOU loss and SSIM loss between the binary black and white image and the camouflaged area image respectively. λ1, λ2 and λ3 represent the loss weight coefficients, and their default values ​​are all 1. i=1, 2, 3, 4 correspond to the four output values ​​of the segmentation network.

[0069] The recognition network part uses the cross entropy loss function L rn For training, the formula is as follows:

[0070] L m = l cross_entropy (predk,label)

[0071] Where predk represents the vector corresponding to the target type output by the recognition network, and label represents the vector corresponding to the actual target type, such as Figure 7 The actual target species shown in the ocean background is a batfish. All target species in the ocean background are [pipefish, batfish, seahorse, …, clownfish] T , then label=[0,1,0,…,0] T , predk=[0.1,0.88,0.01,…,0] T The Adam optimizer was used for training, and the initial hyperparameters were set as batch size batch_size = 32, learning rate learning_rate = 0.0001, total training epochs = 40, and the remaining hyperparameters were set according to the general neural network training rules.

[0072] After each training round, the validation set data is used to test the effect of this round of training, and the internal hyperparameters of the camouflaged target detection model are fine-tuned according to the test results to achieve better training results.

[0073] Step 4, test the disguised target detection model. Export the disguised target detection model trained in step 3 separately, use the test set data to input the disguised target detection model, and calculate the relevant task evaluation indicators based on the output results. For the segmentation network part, use indicators such as MAE, E-MEASURE, F-MEASURE for evaluation; for the recognition network part, calculate the average recognition accuracy of the model. If the relevant indicators meet the actual requirements, the next step can be carried out, otherwise adjust the data set or model hyperparameters and retrain. In this embodiment, it is required that on the COD10K public data set, the segmentation accuracy indicators MAE, E-MEASURE and F-MEASURE reach 0.05, 0.55 and 0.8 respectively, and the recognition accuracy indicator average recognition accuracy reaches more than 85%.

[0074] like Figure 7 The figure shows the detection effect of the camouflaged target detection method proposed in the embodiment of the present application on part of the data of the COD10K public camouflaged target dataset. In the shown ocean and jungle background, the proposed method clearly segments the foreground area of ​​the camouflaged target and correctly identifies the type of the camouflaged target. Fig. 9 It is a qualitative comparison of the segmentation effects of the method of the embodiment of the present application and six relatively new target segmentation technologies (boundary perception network BASNet, gate network GateNet, nested U-type network U2Net, step-by-step feature selection network PFPN, multi-scale feature interaction network MINet, search identification network SINet) on the camouflaged public data set. It can be seen from the figure that the method proposed in the embodiment of the present application has achieved good segmentation effects regardless of whether it is in a variety of complex background environments or for camouflaged objects with large size differences; Table 1 is a quantitative comparison of the method used in the embodiment of the present application and the commonly used target segmentation methods in terms of segmentation indicators. It can be seen from the table below that the method proposed in the embodiment of the present application has achieved a certain degree of leading in multiple public data sets and various indicators.

[0075] Table 1

[0076]

[0077] The present invention provides a method for detecting and identifying camouflaged targets based on a deep neural network. There are many methods and approaches to implement the technical solution. The above is only a specific implementation of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention. All components not specified in this embodiment can be implemented using existing technologies.

Claims

1. A camouflaged target detection and recognition method based on deep neural network, characterized in that: The following steps are involved: Step 1: Constructing a disguised target data set: including collecting disguised target data, image preprocessing, manual annotation and data set division; the data set division refers to dividing the data set into training set data, verification set data and test set data; Step 2: Building a disguised target detection model: The disguised target detection model is used to input a disguised target image and output a disguised target image segmentation and recognition result; Step 3: Train the disguised target detection model: Use the training set data to train the constructed disguised target detection model. During the training process, periodically use the validation set data to conduct preliminary tests on the training results, and adjust the details of the disguised target detection model based on the feedback results of the preliminary tests. Step 4: Test the disguised target detection model: Use the test set data to test the detection effect of the disguised target detection model obtained after training in step 3, and calculate the test evaluation index; The camouflaged target detection model described in step 2 includes a segmentation network and a recognition network: The segmentation network is a fully convolutional neural network, which is used to extract the ROI area of ​​the disguised target image; the input of the segmentation network is the disguised target image, and the output is a binary black and white image containing the disguised target ROI area; the segmentation network includes a front half network, a middle half network and a back half network, and the front half network encodes the disguised target image to obtain a feature map of the disguised target image; Inputting the obtained feature map into the middle part network for feature fusion to obtain a feature fusion result; The feature fusion result is decoded using the second half of the network to obtain a binary black-and-white image containing the camouflaged target ROI area; The recognition network adopts a dual-branch Resnet design to identify the type of the disguised target; the input of the recognition network is the disguised target image and the binary black-and-white image containing the disguised target ROI area obtained by the segmentation network, and the output is the disguised target recognition result; The first half of the segmentation network adopts a U-net structure, which refers to a U-shaped transformation in which the resolution of the obtained feature map changes from large to small and then to large. The left half of the U-net structure is composed of a Resnet-50 backbone network obtained by Imagenet pre-training, which is used to extract and encode multi-scale feature information of the camouflaged target image to obtain four preliminary encoding features. The right half of the U-net structure is composed of four cascaded ordinary convolution blocks, which are used to refine the classification and decoding of the four preliminary encoding features obtained by the left half of the network to obtain four preliminary decoding features; The middle part of the segmentation network is the Receptive Field Block, which splits each of the four outputs of the right half of the U-net structure into four branches, each of which uses convolution kernels of different sizes and convolution methods with different void rates; then, convolution and multiplication are used to fuse the features of the four branches to obtain four feature fusion results; The second half of the segmentation network is an anti-attention module, which inputs four preliminary decoding features obtained by the first half of the segmentation network and four feature fusion results obtained by the middle part to obtain four binary black-and-white images containing the camouflaged target ROI area, including: the preliminary decoding features obtained by the first half of the segmentation network are used to obtain the attention coefficient through the Sigmoid activation function, and the attention coefficient is subtracted from the value 1 to obtain the anti-attention coefficient; the anti-attention coefficient is used to perform the Hadamard product with the feature fusion result obtained by the middle part of the segmentation network, and the calculation result is input into the single-layer volume network to obtain the binary black-and-white image containing the camouflaged target ROI area; The recognition network adopts a dual-branch Resnet structure, directly inputs the camouflaged target image into one Resnet branch, and inputs the binary black-and-white image containing the camouflaged target ROI area obtained by the segmentation network into another Resnet branch after passing through a convolutional network amplification channel; after merging the feature maps output by the two Resnet branches, feature fusion is performed through a continuous three-layer convolutional network, and then input into a two-layer BP neural network to obtain a camouflaged target recognition result.

2. According to claim 1, a method for detecting and identifying camouflaged targets based on a deep neural network is characterized in that: When training the disguised target detection model in step 3, the loss function of the segmentation network is a hybrid loss function, which is expressed by the following formula: Where L n Represents the loss function of the segmentation network part, pred i represents the binary black-and-white image containing the camouflaged target ROI area output by the segmentation network, gt represents the camouflaged area image annotated in step 1.3, and l bce_loss , l iou_loss and l ssim_loss They represent the BCE loss, IOU loss and SSIM loss between the binary black and white image and the disguised area image respectively, λ1, λ2 and λ3 represent the loss weight coefficients, and i=1,2,3,4 corresponds to the four output values ​​of the segmentation network.

3. The method for detecting and identifying camouflaged targets based on a deep neural network according to claim 1, characterized in that: When training the camouflaged target detection model in step 3, the recognition network is trained using the cross entropy loss function L m For training, the formula is as follows: L m =l cross_entropy (predk,label) Among them, predk represents the vector corresponding to the target type output by the recognition network, and label represents the vector corresponding to the actual target type; the training adopts the Adam optimizer, and sets the initial hyperparameters batch size batch_size = 32, learning rate learning_rate = 0.0001, and total training rounds epoch = 40.

4. The method for detecting and identifying camouflaged targets based on a deep neural network according to claim 1, characterized in that: Step 1 includes: Step 1.1, collecting camouflaged target data under complex backgrounds, where the collected data types are static image data or dynamic video data; the complex backgrounds include oceans, jungles, snowy areas, and desert areas; Step 1.2, data preprocessing: the collected camouflaged target data is uniformly cropped into a single-frame color image of size 480×480×3, and then normalized, that is, the image RGB pixel values ​​are normalized to the mean and variance of the ImageNet public dataset to obtain the camouflaged target image; Step 1.3, manual labeling: manually label the actual area of ​​the camouflaged target image, add a label, and mark the target type; Step 1.4, data set division: divide the data set processed in step 1.3 into training set data, validation set data and test set data, and the proportion of the three sets to the total data set are 70%, 15% and 15% respectively.

5. The method for detecting and identifying camouflaged targets based on a deep neural network according to claim 1, characterized in that: The test evaluation indicators described in step 4 include the mean absolute error MAE, enhancement-matching index E-MEASURE and F-MEASURE evaluation indicators for the segmentation network; for the recognition network, the average recognition accuracy of the model is used.

Images