A multi-stage refined camouflage target detection method based on wavelet enhancement recognition

Abstract

The application relates to the technical field of target detection, and discloses a multi-stage refined camouflage target detection method based on wavelet enhancement recognition, which makes up for the defects that the existing method cannot effectively utilize global information, leading to missing of detection details, and cannot process multi-layer information, resulting in feature redundancy problems, so that the method can better complete a position target detection task. First, a visible light image is preprocessed through a pre-training model, a pyramid operation is performed to expand the feature quantity, the amount of calculation is controlled while the amount of information is ensured, part of the initial features are selected for feature enhancement, and the features are screened and identified through a wavelet identification enhancement module; the highest layer features are selected and sent into an asymmetric convolution mamba module to obtain features with global information; then, a feature reverse decoder composed of multiple guide units is used to process the global information features and the identification features to obtain a prediction map, and the model is trained; when a specific detection task is performed, the trained model is used to obtain the prediction map, and the detection task is completed.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of target detection, and in particular to a multi-level refinement camouflage target detection method based on wavelet enhancement recognition. BACKGROUND

[0002] The concept of camouflage target detection is to improve the recognition ability of subtle visual differences. Compared with general target detection and salient target detection, camouflage target detection faces greater challenges due to the seamless integration of camouflage targets with their backgrounds. In recent years, researchers have improved from traditional algorithm methods to deep learning methods.

[0003] Existing camouflage target detection techniques mainly use multi-task methods based on convolutional neural networks. These methods introduce different modules or tasks to improve the ability to detect camouflage targets during the training phase. Generally, when searching for detailed information from coarse to fine, a large amount of features are transferred, and the ability to extract actual useful feature information needs to be strengthened. Moreover, the differences between multiple tasks may have a negative impact on prediction, resulting in feature redundancy problems. Meanwhile, to deal with camouflage target detection in complex environments, some methods have adopted transformer architecture, but they also introduce a large increase in network parameters, resulting in low computational efficiency and making it difficult to deploy on existing devices. However, existing methods cannot guarantee the suppression of missed details while maintaining low computational complexity.

[0004] In the face of the problems of missed details and feature redundancy in existing technologies, there are three main ways to solve this problem: using the self-attention mechanism of transformer to obtain a high receptive field. This method is simple and effective, but the computational cost is too high and it is difficult to deploy in real-world applications. Providing features for camouflage target detection through additional information, such as boundary information. This method requires data to provide support, and also has a large space and time cost. Using a multi-stage method to process features multiple times, relying on data to generate discriminative information. This method needs to improve the use of useful features, but it has a low training cost and does not require additional support from the data set. Currently, it mainly relies on extracting image detail feature information for discrimination.

[0005] However, existing methods usually collect useless redundant feature information in order to ensure computational efficiency, and have weak global attention ability, ignoring the role of global information, so the detection effect needs to be improved. SUMMARY

[0006] The technical problem to be solved by the present application is to provide a multi-level refinement camouflage target detection method based on wavelet enhancement recognition, which can make up for the defects of existing methods that cannot effectively utilize image information to extract effective discriminative features, and ensure low computational complexity for completing the camouflage target detection task while paying attention to global information.

[0007] The technical scheme adopted by the present application to solve the above technical problems is:

[0008] 1. A multi-level refinement camouflage target detection method based on wavelet enhancement recognition, comprising the following steps:

[0009] A: Constructing and training a camouflage target detection model;

[0010] A1: Obtain a sample image and pre-process it, and establish a K-layer pyramid through a Res2Net model; then use a pre-trained Res2Net model to extract features as the first layer features of the pyramid, and generate K groups of features in order through pyramid processing of the features;

[0011] A2: Perform feature processing through a wavelet discrimination enhancement module and an asymmetric convolution mamba module; specifically, based on the obtained pyramid features, use Haar wavelet decomposition to obtain detail discrimination features F wk ; based on the last layer features of the pyramid, use a mamba block and asymmetric convolution to process, respectively, and then aggregate the processed features and input them into a sub-path excitation module to obtain a global feature prediction map

[0012] A3: Reduce the redundant part of the extracted features of the wavelet discrimination enhancement module through a neighborhood aggregation module to obtain a coarse prediction, and reverse the coarse prediction as a reverse mask; set a number of cascaded guide units to form a feature reverse decoder; input the detail discrimination features F wk and the global feature prediction map into the feature reverse decoder to obtain a final prediction map;

[0013] A4: Calculate the weighted binary cross-entropy loss of the last three layers of the pyramid and the weighted intersection over union loss of the overall prediction, respectively, add the two losses to obtain the final loss function of the entire model, and train the camouflage target detection model based on the loss function;

[0014] B: Perform a detection task based on the camouflage target detection model: use the trained camouflage target detection model for actual camouflage target detection tasks.

[0015] Further, the step A1 is specifically as follows:

[0016] Step 11: Pre-processing; obtain a visible light image i to be learned, uniformly set the width and height for each image in the image set I, uniformly convert tensors for RGB images, binary images, and through normalization; group according to the batch size BS;

[0017] Step 12: the pre-processed image i, corresponding to the four-dimensional vector (BS, C, H, W), where BS is the batch size, H is the height, W is the width, and C is the number of channels; the feature is obtained by processing through the pre-trained Res2Net model as the first layer feature f1 of the pyramid;

[0018] Step 13: generating the pyramid feature f based on the first layer feature f1 k ; sequentially generating K groups of features by processing the features through the pyramid layer established by the Res2Net model, each group f k is generated by f k-1 , f k , k∈{1, 2, …, K}, and each layer has a resolution of and the number of channels is C×2 k , and f1 remains unchanged.

[0019] Further, the step A2 is specifically as follows:

[0020] Step 21: the wavelet discrimination enhancement module processes the features as follows: first, the feature f k of the last three layers of the pyramid is converted into 32 through the feature enhancement module, and the four components of the empty convolution are aggregated and reactivated to obtain the enhanced feature f k ′ ; the feature is sent into the Haar wavelet decomposition to obtain four components with a size of , which are a (odd row and odd column), b (odd row and even column), c (even row and odd column), and d (even row and even column); then the features are decomposed into high-frequency information and low-frequency information W k LL :

[0021]

[0022] The high-frequency information is sent into the high-frequency processing with a residual block to obtain the high-frequency processing information HW k :

[0023]

[0024] The low-frequency information is also processed to obtain the low-frequency processing information LW k :

[0025]

[0026] ​Wherein, Res is a batch normalization residual block, AM is a joint attention block combining channel attention and spatial attention, and σ and μ represent the mean and standard deviation of the statistics calculated by the channel, respectively; σ ′ and μ' are affine parameters learned from data, and IN is instance normalization; then the high-frequency and low-frequency information of the sum is mixed by transposed convolution, and the channel information is adjusted by convolution and bilinear interpolation to obtain the detail discrimination feature F with discrimination information wk , which is expressed as follows:

[0027] F wk = FB(TC(HW k +LW k ))

[0028] Wherein, TC represents transposed convolution, and Fb represents a module for adjusting channel information by convolution and bilinear interpolation.

[0029] Step 22: The asymmetric convolution Mamba module processes the feature, which specifically includes the following steps:

[0030] The last layer of the pyramid feature f K is first converted to and which are halved in the number of channels, respectively, and then sent into a Mamba block and an asymmetric convolution decomposed by a 3x3 convolution, respectively, and then aggregated with the last layer feature f K through a 1x1 convolution to obtain a coarse global feature Subsequently, the coarse global feature is obtained by upsampling operation to obtain a prior feature Then, the prior feature is mapped to the same dimension resolution as the latent feature by global average pooling, which will become 1; then activated to obtain a coarse prediction map F prior , which is expressed as follows:

[0031]

[0032] Wherein, C1 and C2 are both 1x1 convolution and contain a normalization function, and δ is a sigmoid gating function for normalization; σ represents a ReLU activation function;

[0033] The coarse prediction map F prior and the coarse global feature are sliced and sent into a sub-path excitation module to output a global feature prediction map Wherein, the sub-path excitation module includes a spatial attention module and a channel attention module, and the process is specifically expressed as follows:

[0034]

[0035] wherein SA denotes a spatial attention module, CA denotes a channel attention module, C3 is a 1x1 convolution and contains a normalization function.

[0036] Further, in the step A3, the feature reverse decoder of each layer is composed of three guide units in series, and the specific feature updating method comprises the following steps:

[0037] Step 31: neighborhood aggregation; first, the neighborhood aggregation module is used to process the high and low frequency information of each layer to generate a reverse mask:

[0038]

[0039] wherein, denotes the k-th layer of pyramid feature, the reverse mask of the i-th guide unit, i∈{1,2,3}; k is specifically the last three layers of the pyramid; NCD denotes neighborhood aggregation, Sampling denotes sampling, and E denotes an all-one matrix, denotes the inverse operation used to subtract the matrix E;

[0040] Step 32: set the group number g i , g i ∈{32,8,1}, one guide unit is used for each group, and the group g i is set wk , wherein j∈{1,……,g i}, so as to obtain the slice feature When i∈{1,3}, the guide unit processing is summarized as follows:

[0041]

[0042]

[0043] wherein Split denotes grouping, Cat denotes aggregation; when i=2, an attention module and a random inactivation dropout are introduced to suppress redundant features, and the slicing operation is as follows:

[0044]

[0045] wherein conv denotes convolution operation, PA and CA denote position attention module and channel attention module respectively, and drop denotes dropout operation with a coefficient of 0.1;

[0046] Through the aggregation formula Cat, three refined guide groups are obtained

[0047] Then, a reverse conversion process is performed to obtain a finer layer of prediction mask The specific process is as follows:

[0048]

[0049] where C s denotes the operation of 3x3 kernel convolution and batch normalization on the slice, and Score represents the operation of converting data into learnable weight parameters through convolution;

[0050] Then, a more accurate prediction is obtained in the following form:

[0051]

[0052] where sampling represents a sampling operation opposite to the neighborhood aggregation sampling operation, k∈{K-2,K-1,K}, K represents the last pyramid layer; when k is the last layer, Pred k+1 The input of Pred According to the above formula, the accurate prediction map Pred k is obtained in sequence. k The accurate prediction map Pred y when k=K-2 is taken as the final prediction camouflage target detection prediction map.

[0053] Further, the weighted binary cross-entropy loss is represented as follows:

[0054]

[0055] where y is the true label, 0 or 1, represented by our actual prediction map, is the positive class probability predicted by the model; w y and w 1-y are the weights for the positive class and the negative class, respectively.

[0056] The weighted intersection over union loss is represented as follows:

[0057]

[0058] For which and B represent the real target bounding box and the predicted bounding box of the camouflage target, respectively, and w is the weight of the bounding box.

[0059] Further, the final loss function of the model is represented as follows:

[0060]

[0061] where α is a weight parameter, and GT represents the image ground truth.

[0062] Further, the weight parameter α is obtained based on tanh(x+1) according to the training step number x, specifically, when k is the last layer, When k is the last 2, 3 layers,

[0063] Further, the pyramid layer number K=5.

[0064] Further, the sampling operation of the neighborhood aggregation is specifically that when k=5, ×4 up-sampling is performed, and when k is in the set {3, 4}, ×2 down-sampling is performed.

[0065] Further, the enhanced feature f k ′ Specifically, the following method is used:

[0066] For the feature set f k ,k∈{3,4,5}, a residual layer component b i ∈{1,2,3,4} feature enhancement module is introduced, the expansion rate d of the four parallel residual layer components is ∈{1,3,5,7}, the channel size is reduced to 32 by 1×1 convolution in the first layer of each component, and for the residual layer, an (2i-1)×(2i-1) convolution layer and a 3×3 convolution layer with expansion rate d are included, the channel number is controlled at 32, the component results are added and sent to the ReLU function activation to finally obtain the enhanced feature f k ′ ,k∈{3,4,5}.

[0067] The beneficial effects of the present application are:

[0068] 1) The advantages and disadvantages of convolutional neural networks and transformer methods are considered, the mamba and convolutional network are combined, the multi-level refined feature information is obtained, the useful discrimination information is obtained, the long-range dependence of information is strengthened under the premise of low calculation amount, and the receptive field is improved. Thus, the effect of camouflage target detection in the case of missing details is improved.

[0069] 2) The time-frequency positioning characteristics of the wavelet are considered to enhance the image and reduce the noise, and at the same time, in the case of multi-layer information features, a feature inverse decoder is used for decoding operation, and in the form of a mask, the feature redundancy and noise are reduced, so that the non-target is regarded as a camouflage target in the detection, and the accuracy of the camouflage target detection under the problems of missing details and feature redundancy is further improved. BRIEF DESCRIPTION OF DRAWINGS

[0070] Figure 1 The overall flowchart of the camouflage target detection method in the present application is shown in the figure;

[0071] Figure 2 The principle diagram of the wavelet discrimination enhancement of the present application is shown in the figure;

[0072] Figure 3Schematic diagram of asymmetric convolution Mamba for the present application;

[0073] Figure 4 Schematic diagram of guiding unit in feature reverse decoder for the present application;

[0074] Figure 5 Process schematic diagram of camouflage target detection for the present application. DETAILED DESCRIPTION

[0075] The present application aims to propose a multi-level refinement camouflage target detection method based on wavelet enhancement recognition, which makes up for the defects that the existing method cannot effectively utilize global information for complex image detection and cannot process multi-layer information to produce feature redundancy, so that it can better complete the camouflage target detection task under the two defect challenges. The overall implementation process is as shown in the figure: Figure 1

[0076] S1, the pre-trained model is used to process the visible light image to be learned, and the image is converted into multi-layer initial features through the pyramid structure;

[0077] S2.1, the last three layers of initial features are sent to the feature enhancement module for processing, and then updated by the wavelet discrimination enhancement module after wavelet decomposition;

[0078] S2.2, the initial features of the last layer are updated by the asymmetric convolution Mamba module to reserve global information;

[0079] S3, the features are processed by the feature reverse decoder to obtain the prediction map, each decoder contains 3 sequentially connected guiding units, and the wavelet discrimination enhancement module of each layer updates the features by taking the prediction of the previous layer as a mask, and the features updated by the asymmetric convolution Mamba are taken as a mask for the first time;

[0080] S4, the weighted intersection over union loss and the weighted binary cross entropy loss are calculated respectively, and the two kinds of losses are combined and then the Adam optimizer is used to train the camouflage target detection model.

[0081] S5, the specific camouflage target detection task is executed, and the trained model is used to obtain the final prediction map when detecting the unlearned image.

[0082] Embodiment:

[0083] Taking a visible light image i to be learned with a height of H, a width of W and a channel number of 3 as an example, the present application scheme is described in detail, and the specific implementation process is as follows:

[0084] ​Step 1, first, the sample image is preprocessed; the data preprocessing operation specifically includes three sub-steps of data preprocessing, pre-training model feature extraction, and pyramid model processing features:

[0085] Step 11, in the form of a tensor, each image in the image set I is uniformly set to 384*384 by changing the width and height, and the RGB image and the binary image are uniformly converted into tensors by normalization; grouping is performed according to the batch size BS.

[0086] Step 12, the pre-trained image obtained in step 11 is processed by Res2Net pre-trained on a large visual database imagenet to extract features:

[0087] For a specific image i(3,H,W), its corresponding four-dimensional vector is (BS,C,H,W), set H and W to 384, C to 32, first process through the pre-trained Res2Net model, and obtain parameters such as (BS,64,96,96) features through the activation function and the maximum pooling layer in the pre-trained model, and the features are taken as the first layer f1 of the pyramid.

[0088] Step 13, generate pyramid features f k based on existing features f1

[0089] Process the features in sequence through the pyramid layer established by the Res2Net model to generate 5 groups of features, each group f k is generated from f k-1 , f k , k∈{1,2,3,4,5}, and f1 is taken as the reference, each layer has a resolution of and a channel number of C×2 k , and f1 remains unchanged.

[0090] Step 2, process the features through the wavelet discrimination enhancement module and the asymmetric convolution mamba module respectively:

[0091] Among them, different layers k determine the size of the four-dimensional vector resolution, considering that too large resolution in the lower layer will increase the calculation cost, so k usually takes 3, 4, 5 in the wavelet discrimination enhancement module. The asymmetric convolution mamba module only takes the last layer k=5 considering global information;

[0092] Step 21, as shown in Figure 2 , the wavelet discrimination enhancement module processes the features as follows: first, the feature enhancement module converts the channel number C to 32, and after aggregating and reactivating the components with 4 hollow convolutions, an enhanced feature f is obtained, which is similar to k ′ . Specifically, for the feature set fk , k e {3, 4, 5}, introduce a residual layer component b with 4 different expansion rates i , k e {3, 4, 5}. The feature enhancement module, 4 parallel residual layer components with expansion rate d e {1, 3, 5, 7}, the first layer of each component is a 1x1 convolution to reduce the channel size to 32, for the residual layer which includes a (2i-1)x(2i-1) convolution layer and a 3x3 convolution layer with expansion rate d, the channel size is controlled at 32, the component results are summed and sent to the ReLU function activation to obtain the enhanced features f k ′ , k e {3, 4, 5}.

[0093] The feature is sent to the Haar wavelet decomposition, which contains two filters, row filter and column filter. After passing through the row filter, the image is decomposed by height dimension and divided by two, and two row odd and even features of the same size can be obtained After passing through the column filter, the row odd and even features are decomposed by column, and four components of size a (row odd, column odd), b (row odd, column even), c (row even, column odd), d (row even, column even) are obtained, and then the features are decomposed into high frequency information and low frequency information

[0094]

[0095] LL, LH, HL, HH represent low frequency information and horizontal, vertical, diagonal high frequency information respectively. The high frequency information is sent to the high frequency processing with residual block to obtain high frequency processing information HW k , for example, when k = 5:

[0096]

[0097] Similarly, the low frequency information is also processed to obtain low frequency processing information LW5:

[0098]

[0099] Where Res is the batch normalization residual block (BN), AM is the joint attention block combining channel attention and spatial attention, and σ and μ represent the mean and standard deviation of the statistical quantity calculated by channel. σ ′and μ' are affine parameters learned from data, IN is instance normalization. The high-frequency and low-frequency information of the sum are mixed by transposed convolution (TC), two convolution layers and one up-sampling are performed by Fb module, a jump connection is added on the last feature, and bilinear interpolation is used to adjust the feature channel information:

[0100] F wk = Fb(TC(HW k +LW k ))

[0101] The detail discrimination feature F wk with discrimination information is obtained here w5 ; for example, when k = 5, F prior its four-dimensional information is adjusted to

[0102] Step 22, as shown in Figure 3 , the asymmetric convolution Mamba module processes the feature, which specifically includes the following steps:

[0103] The feature f5 obtained in step 13 is first converted to and , both of which are halved in the number of channels, and are respectively sent into a Mamba block and an asymmetric convolution decomposed by a 3x3 convolution, wherein the Mamba block is from the base layer in the visual state space model; then the coarse global feature is obtained by aggregating f5 through 1x1 convolution, which is a four-dimensional vector (BS, 2048, 12, 12), and then the coarse global feature is obtained by channel mapping and up-sampling operation with a scale of 32, and the prior feature is obtained, which is converted to a resolution of 384, and then the vector is mapped to the same dimension as the latent feature by global average pooling, and the resolution will become 1, and the specific process can be represented as follows:

[0104]

[0105] GAP(X) is the formula of global average pooling, X is a variable, and here we will process the prior feature ;

[0106]

[0107] , wherein C1, C2, and C3 are all 1x1 convolution and contain a normalization function, and δ is a sigmoid gating function for normalization; σ represents a ReLU activation function, and after activation, the coarse prediction map F prior is obtained, which is (BS, 1, 12, 12); prior and the coarse global feature The slice is sent to the sub-path excitation (SDE) module, and the accurate output contains global feature information of the image, i.e. global feature prediction map Wherein the SDE module includes a spatial attention SA module and a channel attention CA module, and is specifically expressed as follows:

[0108]

[0109] Step 3, obtain the detail discrimination feature F by step 2 wk And the global feature prediction map Then, a feature inversion decoder is used for decoding to obtain a final prediction map, and the feature inversion decoder of each layer is composed of three guide units connected in series, and the specific updating method includes the following steps:

[0110] Step 31, neighborhood aggregation. The three-layer detail discrimination feature F obtained in step 21 w3 ,F w4 ,F w5 With the same source of the initial feature f1 in step 12, in order to reduce feature redundancy, the feature information capable of cross-layer connection is obtained by convolution and up-sampling method. As shown in Figure 4 After the neighborhood aggregation NCD module, the high and low frequency information TC(HW k +LW k ) of each layer without Fb module processing is used to generate the first reverse mask Wherein k=5 represents the fifth group of features, and i=1 represents the mask in the first guide unit of the sequence:

[0111]

[0112] Wherein, TC refers to transpose convolution, σ is a sigmoid function, Sampling sampling refers to ×4 up-sampling when k=5, and ×2 down-sampling when k is in the set {3, 4}. The matrix E is a full one matrix, Indicates the inverse operation of the matrix E used for subtraction.

[0113] Step 32, set the number of groups g i ,i∈{1,2,3},g i ∈{32,8,1},each group uses a guide unit, and j is determined by dividing the total number of feature channels by g i For example Set k=5 and i=1, set the group g i Subsequently, the feature F wk is split, where j∈{1,……,g i}, so as to obtain the slice feature, and the conventional guide unit processing is summarized as follows:

[0114]

[0115]

[0116] In the above operations, i∈{1,3}, where Split represents grouping and Cat represents aggregation. When processing the middle group, i=2, we add an attention module and introduce dropout to suppress redundant features. Accordingly, the slicing operation is as follows, taking g2=8 as an example:

[0117]

[0118] Where j∈{1,…,8}, conv represents the convolution operation, PA and CA represent the position attention module and channel attention module respectively. drop represents the dropout operation with a coefficient of 0.1. Then, the aggregation formula Cat is used to aggregate the three refined guidance groups with different degrees of detail. Each of these features is then subjected to a reverse transformation process to obtain a finer layer of prediction masks, starting with the first reverse mask For example, the input is This is obtained by aggregating Cat and grouping Split Obtained through intermediate guidance Substitute it into the following formula:

[0119]

[0120] Among them C s Refers to the operation after performing 3×3 kernel convolution on the slice and performing batch normalization. The "Score" operation converts the data into learnable weight parameters through convolution. Then we get Obtaining a more precise forecast then takes the form:

[0121]

[0122] Where “sampling” refers to the sampling operation opposite to the sampling operation in neighborhood aggregation NCD, k∈{3,4,5}. Specifically, when k=5, Pred k+1 The input is You can get

[0123]

[0124] Similarly, we obtain Pred4 and Pred3 in sequence according to the above formula, and use Pred3 as the final predicted camouflaged target detection prediction map.

[0125] Step 4, the final predicted camouflage target detection prediction map is obtained by decoding in step 3, in this step, the weighted binary cross entropy and weighted intersection over union loss values will be calculated respectively, and the model parameters matrix will be updated through the Adam optimizer training. Specifically, it includes the following steps:

[0126] Step 41, the weighted binary cross entropy loss function is calculated by the following formula:

[0127]

[0128] Where y is the true label (0 or 1) represented by our actual prediction map, is the positive class probability predicted by the model, that is, the probability of the binary true value image taking 1 part in camouflage target detection, where w y and w 1-y are the weights for positive and negative classes respectively. is written as

[0129] Step 42, calculate the weighted intersection over union loss function:

[0130]

[0131] For which and B represent the real target bounding box and the predicted bounding box of the camouflage target respectively, calculate their intersection area and union area, the intersection area is the area of the overlapping part of the predicted box and the real box, and the union area is the total area of the predicted box and the real box. Divide the intersection area by the union area to get the IoU (intersection over union), for each bounding box, assign a weight w according to its class, then calculate the weighted IoU loss;

[0132] A53, combine the weighted binary cross entropy loss function and the weighted intersection over union loss function L wIoU to obtain the final loss L t of the model, and train the camouflage target detection model by minimizing the final loss L t through the Adam optimizer;

[0133]

[0134] Where GT represents the image ground truth; here we add the weight parameter α, according to the training step number, based on tanh(x+1), x increases by 1 every 100 steps, when k = 3, 4, when k = 5,

[0135] Step 5, after the model training is completed through step 4, the model can be used for actual camouflage target detection tasks. For example, for the image G to be detected, the hidden camouflage target therein is to be predicted. As shown in FIG. 6, the trained model is selected, and the last Pred3 is obtained through steps 1, 2, and 3 to obtain a prediction map as a result output. Figure 5

[0136] Although the present application has been described with reference to the embodiments thereof, it is to be understood that the embodiments are intended to be illustrative only and the application is not limited to those embodiments. Numerous modifications, equivalents, and alternatives are apparent in light of the disclosure herein without departing from the spirit and scope of the present application.​

Claims

1. A multi-level refinement camouflage target detection method based on wavelet enhanced recognition, characterized in that: The following steps are involved: A: Build and train a camouflaged object detection model; Step A1: Obtain a sample image and preprocess it, then use the Res2Net model to build a K-layer pyramid. Then, use the pre-trained Res2Net model to extract features as the first layer of the pyramid. K groups of features are generated sequentially through the pyramid processing features. Step A2: Feature processing is performed using the wavelet discrimination enhancement module and the asymmetric convolution Mamba module respectively. Detail discrimination features are obtained using Haar wavelet decomposition based on the obtained pyramid features. The features of the last layer of the pyramid are processed using Mamba blocks and asymmetric convolution respectively. The processed features are then aggregated and input into the subpath excitation module to obtain a global feature prediction map. Step A3: Using the neighborhood aggregation module to reduce the redundant parts of the features extracted by the wavelet discrimination enhancement module to obtain a rough prediction, and inverting the rough prediction as a reverse mask; forming a feature inversion decoder by a set number of cascaded guide units; The detail discrimination features and the global feature prediction map are decoded using a feature inversion decoder to obtain the final prediction map; Step A4: Calculate the weighted binary cross entropy loss of the last three layers of the pyramid and the weighted intersection-over-union loss of the overall prediction, add the two losses together to obtain the final loss function of the entire model, and train the disguised object detection model based on the loss function; B: Perform detection tasks based on the disguised target detection model: Use the trained disguised target detection model for the actual disguised target detection task.

2. The multi-level refinement camouflage target detection method based on wavelet enhanced recognition according to claim 1 is characterized in that: The step A1 is specifically as follows: Step 11: Preprocessing: Get the visible light image i to be learned, uniformly set the width and height for each image i in the image set I, and uniformly convert the tensor of RGB images and binary images through normalization; group them according to the batch size BS; Step 12: The preprocessed image i, whose corresponding four-dimensional vector is (BS, C, H, W), where the batch size is BS, the height is H, the width is W, and the number of channels is C; is processed by the pre-trained Res2Net model to obtain the features as the first layer feature f1 of the pyramid; Step 13: Generate pyramid features f based on the first layer features f1 k ; K groups of features are generated sequentially by processing the features at the pyramid layer established by the Res2Net model, f k ,k∈{1,2,…,K}, taking f1 as the benchmark, the resolution of each layer is The number of channels is C×2 k , f1 remains unchanged.

3. The multi-level refinement camouflage target detection method based on wavelet enhanced recognition according to claim 2 is characterized in that: The step A2 is specifically as follows: Step 21: The wavelet discrimination enhancement module processes the features as follows: First, the feature enhancement module is used to transform the features f of the last three layers of the pyramid into k The number of channels C is converted to 32, and the four components of the dilated convolution are aggregated and reactivated to obtain the following The enhanced feature f k ′ ; Send the feature into the Haar wavelet decomposition and transform it into a size of The four components are row odd and column odd a, row odd and column even b, row even and column odd c, row even and column even d; then the feature is decomposed into high-frequency information by the following formula With low-frequency information High-frequency information Send it to the high frequency processing with residual block to obtain high frequency processing information HW k : Low-frequency information Also processed to obtain low frequency processing information LW k : Among them, Res is the batch normalized residual block, AM is the joint attention block that combines channel attention and spatial attention, σ and μ represent the mean and standard deviation of the statistics calculated by the channel respectively; σ ′ and μ′ are affine parameters learned from the data, IN is instance normalization; then the high-frequency and low-frequency information are mixed and summed by transposing convolution, and then convolution is performed and bilinear interpolation is used to adjust the channel information to obtain the detail discriminative feature Fw with discriminative information k , which is expressed as follows: Fw k =Fb(TC(HW k +LW k )) Among them, TC represents transposed convolution, and Fb represents a module that adjusts channel information by performing convolution and bilinear interpolation; Step 22: The asymmetric convolutional mamba module processes features in the following steps: The last layer of the pyramid feature f K First, transform it into and The number of channels is halved and fed into a Mamba block and an asymmetric convolution decomposed by a 3×3 convolution, and then combined with the last layer feature f by a 1×1 convolution. K Aggregate to obtain coarse global features Then the coarse global features Obtain prior features through upsampling operations Then the prior features are pooled by global average pooling GAP Mapped to the same dimension as the potential feature, the resolution will become 1; then the activation is used to obtain the coarse prediction map F prior , which is expressed as follows: Where C1 and C2 are both 1×1 convolutions and contain a normalization function, and δ is the sigmoid gating function used for normalization; σ represents the ReLU activation function; The rough prediction map F prior and coarse global features Slice and send to the subpath excitation module to output the global feature prediction map The subpath excitation module includes the spatial attention module and the channel attention module. The specific process is as follows: Among them, SA represents the spatial attention module, CA represents the channel attention module, and C3 is a 1×1 convolution and contains a normalization function.

4. The multi-level refinement camouflage target detection method based on wavelet enhanced recognition according to claim 3 is characterized in that: In step A3, the feature inversion decoder of each layer is composed of three guide units connected in series, and its specific feature updating method includes the following steps: Step 31: Neighborhood aggregation: First, the neighborhood aggregation module is used to process the high and low frequency information of each layer to generate a reverse mask: in, represents the k-th layer pyramid feature, the reverse mask of the i-th guide unit, i∈{1,2,3}; k specifically refers to the last three layers of the pyramid; NCD represents neighborhood aggregation, Sampling represents sampling, and E represents the all-one matrix. The inverse operation used to subtract the matrix E; Step 32: Set the number of groups g i , g i ∈{32,8,1}, each group uses a guide unit, set the group g i Then split the feature Fw k , where j∈{1,……,g i }, thus obtaining the slice feature , the boot unit processing is outlined as follows: Where Split represents grouping and Cat represents aggregation. When i=2, the attention module and random inactivation dropout are introduced to suppress redundant features. The slicing operation is expressed as follows: Among them, conv represents the convolution operation, PA and CA represent the position attention module and channel attention module respectively; drop represents the dropout operation; Through the aggregation formula Cat aggregation, three refined guidance groups are obtained The reverse conversion process is then performed to obtain a finer layer of prediction mask The specific process is as follows: Among them C s Refers to the operation of performing 3×3 kernel convolution on the slice and performing batch normalization. Score represents the conversion of data into learnable weight parameters through convolution; A more precise forecast is then obtained using: Among them, sampling represents the sampling operation opposite to the neighborhood aggregation sampling operation, k∈{K-2,K-1,K}, K represents the last pyramid layer; when k is the last layer, Pred l+1 The input is Get the accurate prediction graph Pred according to the above formula sequence l , the precise prediction graph Pred when k=K-2 l Prediction map for disguised object detection as the final prediction.

5. The multi-level refinement camouflage target detection method based on wavelet enhanced recognition according to claim 4 is characterized in that: The weighted binary cross entropy loss is expressed as follows: Among them, y is the true label, 0 or 1, which is represented by our actual prediction map. is the positive class probability predicted by the model, w y and w 1-y are the weights for positive and negative classes respectively; The weighted intersection-over-union loss is expressed as follows: For which and B represent the disguised target real target bounding box and predicted box respectively, and w is the weight of the bounding box.

6. The multi-level refinement camouflage target detection method based on wavelet enhanced recognition according to claim 5 is characterized in that: The final loss function of the model is expressed as follows: Among them, α is the weight parameter and GT represents the true value of the image.

7. The multi-level refinement camouflage target detection method based on wavelet enhanced recognition according to claim 6 is characterized in that: The weight parameter α is obtained based on tanh(x+1) according to the number of training steps x. Specifically, when k is the last layer, When k is the last 2 or 3 layers, 8. The multi-level refinement camouflage target detection method based on wavelet enhanced recognition according to claim 7 is characterized in that: The number of layers of the pyramid is K=5.

9. The multi-level refinement camouflage target detection method based on wavelet enhanced recognition according to claim 8 is characterized in that: The sampling operation of the neighborhood aggregation is specifically to perform ×4 upsampling when k=5, and to perform ×2 downsampling when k is in the set {3, 4}.

10. The multi-level refinement camouflage target detection method based on wavelet enhanced recognition according to claim 9 is characterized in that: The enhanced feature f k ′ Specifically obtained through the following methods: For the feature set f k ,k∈{3,4,5}, introduce a residual layer component b with 4 different expansion rates i ∈{1,2,3,4} feature enhancement module, the expansion rate d∈{1,3,5,7} of the four parallel residual layer components, the first layer of each component is a 1×1 convolution to reduce the number of channels to 32, the residual layer includes a (2i-1)×(2i-1) convolution layer and a 3×3 convolution layer with expansion rate d, the number of channels is controlled at 32, the component results are summed and sent to the ReLU function for activation to finally obtain the enhanced feature f k ′ ,k∈{3,4,5}.

Images