Camouflage target detection method based on cascade frequency domain perception and refined feature guidance

Abstract

The invention relates to the technical field of computer vision, and discloses a camouflage target detection method based on cascade frequency domain perception and refinement feature guidance, comprising the following steps: a computer device obtains a camouflage image and extracts a multi-scale feature map; performing frequency domain separation by using a frequency domain sensing module, generating low-frequency and high-frequency characteristic sub-bands, and performing cross-hierarchy aggregation from bottom to top through a cascaded frequency domain sensor to obtain frequency fusion characteristics; receiving the frequency fusion features from top to bottom by using an aggregation guide decoding unit, performing guide refinement in combination with the deep decoding features, adapting to irregular edges by using a variable kernel convolution strategy in the refinement process, and generating a prediction feature map in combination with a partial convolution strategy; and outputting the predicted feature map as a detection result. According to the invention, through frequency domain information complementation and dynamic convolution refinement, the problem of difficult detection caused by highly similar texture of the camouflage target and the background is solved, and the positioning accuracy and edge segmentation integrity of the camouflage target are improved.

Description

Technical Field

[0001] This invention relates to the field of computer vision technology, specifically to a method for detecting camouflaged targets based on cascaded frequency domain perception and refined feature guidance. Background Technology

[0002] Camouflaged target detection is an important topic in the field of computer vision. Its core task is to identify and segment targets with camouflage characteristics in images that are highly integrated with the background. This technology has important application value in many fields such as biodiversity conservation, field search and rescue, military reconnaissance, and medical image analysis. The characteristic of camouflaged targets is that the target itself achieves a high degree of integration with the background by simulating the color, texture and shape of the surrounding environment, resulting in almost no clear boundary and obvious visual difference between the target and the background.

[0003] Existing methods for detecting camouflaged targets mainly rely on deep convolutional neural networks to extract spatial domain features of images. These methods capture color, texture, and local structural information of images by stacking convolutional layers and combine attention mechanisms to enhance the feature representation of salient regions. However, when the texture features of the camouflaged target are extremely similar to those of the background, network models that rely solely on spatial domain feature extraction will confuse the information of the target with that of the background, resulting in insufficient feature extraction and thus missed detection or misidentification of the camouflaged target.

[0004] Furthermore, existing detection methods face challenges when dealing with the irregular edges of camouflaged targets. Camouflaged targets, especially organisms in nature, have irregular contours and blurred boundaries. Standard convolution operations use fixed rectangular sampling kernels. This fixed geometry makes it difficult to classify and locate edge pixels when processing smooth, continuous, blurred boundaries. This directly leads to coarse and inaccurate segmentation results at the target edges, affecting the overall accuracy of camouflaged target detection. Therefore, this invention proposes a camouflaged target detection method based on cascaded frequency domain sensing and refined feature guidance to address the shortcomings of existing technologies. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a camouflage target detection method based on cascaded frequency domain sensing and refined feature guidance. This method solves the problems in existing camouflage target detection tasks, such as insufficient feature extraction due to the high similarity between the camouflage target and the background texture, and low positioning accuracy due to the blurry and irregular edges of the target.

[0006] To achieve the above objectives, the present invention provides a camouflaged target detection method based on cascaded frequency domain sensing and refined feature guidance, comprising the following steps: The computer equipment acquires the camouflaged image to be detected and inputs the camouflaged image into the feature extraction module to generate feature maps at four different scales; The frequency domain sensing module is used to perform frequency domain feature separation on feature maps of four different scales to generate corresponding low-frequency feature bands and high-frequency feature bands. The frequency doubling convolution strategy is then used to process the low-frequency feature bands and high-frequency feature bands respectively to obtain low-frequency processed features and high-frequency processed features. Using the cascaded frequency domain perceptron built in the frequency domain sensing module, low-frequency processing features and high-frequency processing features are aggregated across levels in a bottom-up order to generate four levels of frequency fusion features. Using the aggregation-guided decoding unit built in the decoding and refinement module, frequency fusion features are received in a top-down order, and combined with the decoded features from deep layers for guided refinement. During the refinement process, the features are processed using a variable kernel convolution strategy and a partial convolution strategy to generate four levels of predicted feature maps. The predicted feature map of the first level is output as the camouflaged target detection result, and the loss calculation module is used to calculate the loss value between the predicted feature maps of all levels and the real label map. The network parameters are updated through the backpropagation algorithm.

[0007] Preferably, the process of performing frequency domain feature separation on feature maps of four different scales using the frequency domain sensing module includes: By performing convolution operations between the low-pass filter weight matrix in the depth wavelet transform unit and the feature map, and superimposing bias terms, a low-frequency feature band is generated. The low-frequency feature band preserves the background and target structure information that changes gradually in the camouflaged image. By performing convolution operations between the high-pass filter weight matrix in the depth wavelet transform unit and the feature map, and superimposing bias terms, a high-frequency feature band is generated. The high-frequency feature band is then used to extract information on the boundaries of drastically changing objects and subtle texture differences in the camouflaged image. The low-pass filter weight matrix and the high-pass filter weight matrix are set as learnable parameters during network training to adapt to the frequency distribution characteristics of different camouflage scenarios.

[0008] Preferably, the process of cross-level aggregation of low-frequency and high-frequency processing features in a bottom-up order includes: For the first level, the frequency fusion unit receives the low-frequency processing features and the high-frequency processing features of the first level to generate the frequency fusion features of the first level. For the second to fourth levels, the frequency fusion unit is used to receive the low-frequency processing characteristics of the current level, the high-frequency processing characteristics of the current level, and the frequency fusion characteristics from the previous level. The frequency fusion unit takes the frequency fusion features from the previous level as prior information and fuses them with the low-frequency processing features and high-frequency processing features of the current level to generate the frequency fusion features of the current level, thereby achieving effective complementarity of multi-scale frequency domain information.

[0009] Preferably, the specific process by which the frequency fusion unit uses the frequency fusion features from the previous level as prior information and fuses them with the low-frequency processing features and high-frequency processing features of the current level includes: The frequency fusion features from the previous level are downsampled and channel adjusted, and then normalized and activated to generate cross-level intermediate features. The intermediate features across levels are concatenated with the high-frequency processing features of the current level in the channel dimension and input into the constructed high-frequency enhancement unit. Features are preserved through the residual connection structure, and spatial weights and channel weights are calculated using the collaborative attention mechanism to generate enhanced high-frequency features. The cross-level intermediate features and the low-frequency processing features of the current level are concatenated in the channel dimension and input into the constructed low-frequency enhancement unit. The residual processing of instance normalization constraints, the position normalization processing and the collaborative attention mechanism processing are executed in sequence to generate enhanced low-frequency features. The enhanced high-frequency features, cross-level intermediate features, and enhanced low-frequency features are added element-wise and then fused through a convolutional layer to obtain the frequency fusion features of the current level.

[0010] Preferably, the process of using the aggregation-guided decoding unit constructed in the decoding and refinement module to receive frequency fusion features in a top-down order and combine them with deep decoding features for guided refinement includes: For the fourth level, the frequency fusion features of the fourth level are directly decoded to generate the predicted feature map of the fourth level. For the first to third levels, the aggregation-guided decoding unit receives the frequency fusion features of the current level and the predicted feature map from the next level. An upsampling operation is performed on the predicted feature map from the next level to make the size of the predicted feature map from the next level consistent with the frequency fusion feature of the current level. The predicted feature map from the next level and the frequency fusion feature of the current level are concatenated along the channel dimension and then passed through a convolutional layer to generate the initial fusion feature.

[0011] Preferably, the process of using a variable kernel convolution strategy in the refinement process includes: inputting the initial fused features into a variable kernel convolution layer; The variable kernel convolutional layer dynamically adjusts the sampling position of the convolutional kernel according to the spatial distribution of the initial fused features, generating sampling coordinates of arbitrary shapes to match the edge morphology of the camouflaged target; By using variable kernel convolutional layers to map the initial fused features into parallel first-branch and second-branch features, the limitations of traditional fixed convolutional kernels in extracting features from irregularly camouflaged targets are overcome.

[0012] Preferably, the process of using partial convolution strategies to process features during the refinement process includes: Perform partial convolution operations on the features of the first branch and the features of the second branch respectively; The feature channel is divided into a first sub-part that participates in the calculation and a second sub-part that remains unchanged; Convolution operations are performed only on the first sub-part, the result is re-concatenated with the second sub-part, and then fused through pointwise convolution to generate the first and second thinning features respectively; The first and second refined features are added element-wise and an upsampling operation is performed to generate the predicted feature map of the current level, thereby reducing computational redundancy and improving feature representation efficiency.

[0013] Preferably, the process of calculating the loss value between the predicted feature maps and the ground truth label maps at all levels using the loss calculation module includes: Construct a hybrid loss function, which consists of a weighted binary cross-entropy loss function and a weighted intersection-over-union (IoU) loss function; The predicted feature maps from the first to the fourth level are compared with the real label maps, where the real label maps are binarized images marked with the disguised target regions. For each level, the difference between the predicted feature map and the real label map in pixel classification is calculated using the weighted binary cross-entropy loss function, and the difference between the predicted feature map and the real label map in target region overlap is calculated using the weighted intersection-union loss function. The total loss value is obtained by summing the calculation results of all levels, thus achieving supervision of network training.

[0014] Preferably, the weighted binary cross-entropy loss function assigns weights based on the classification difficulty of pixels. Pixels located at the edge of a camouflaged target or in a region with blurred texture are assigned greater weights than pixels located in a flat background region, thereby guiding the network model to pay more attention to edge pixels that are difficult to distinguish.

[0015] Preferably, the feature extraction module generates feature maps using a backbone network based on a hybrid architecture, the backbone network including convolutional neural network layers and Transformer layers; In the stage of generating the first-level feature map and the second-level feature map, convolutional neural network layers are used to extract local texture and edge geometry information; In the stages of generating the third-level and fourth-level feature maps, the self-attention mechanism in the Transformer layer is used to capture global contextual information and semantic category information, locate the overall region of the camouflaged target, and achieve complementarity between local details and global semantics.

[0016] This invention provides a camouflaged target detection method based on cascaded frequency domain sensing and refined feature guidance. It has the following beneficial effects: 1. This invention utilizes a frequency domain perception module to perform frequency domain feature separation on the feature map, decomposing the camouflaged image into low-frequency feature bands representing the main structure and high-frequency feature bands representing edge texture. It also combines a cascaded frequency domain perceptron for cross-level aggregation. The camouflaged target detection method based on cascaded frequency domain perception and refined feature guidance can use frequency domain information as a supplement to capture the subtle frequency differences between the camouflaged target and the background. This solves the problem of insufficient target feature extraction and missed detection caused by traditional methods relying only on spatial domain features in camouflaged scenes where visual textures and backgrounds are highly similar. This improves the ability to detect highly concealed targets.

[0017] 2. This invention constructs an aggregation-guided decoding unit in the decoding and refinement module, and integrates a variable kernel convolution strategy and a partial convolution strategy during the refinement process. The variable kernel convolution strategy dynamically adjusts the sampling position of the convolution kernel according to the spatial distribution of features, thereby generating sampling coordinates that can adaptively match the irregular shape of the camouflaged target. The camouflaged target detection method based on cascaded frequency domain perception and refined feature guidance breaks through the limitation of traditional fixed convolution kernels' poor adaptability to geometric deformation, and can fit the complex biological edges of the camouflaged target. At the same time, the partial convolution strategy reduces redundant calculations, improving the precision of edge segmentation while ensuring the efficiency of feature expression.

[0018] 3. This invention employs a hybrid backbone network comprising convolutional neural network layers and Transformer layers, trained using a hybrid loss function consisting of a weighted binary cross-entropy loss function and a weighted intersection-over-union (IoU) loss function. The hybrid backbone network combines the ability of convolutional neural networks to extract local texture details with the ability of Transformers to capture global semantic context, achieving feature complementarity. The weighted hybrid loss function, by increasing the weights of hard-to-classify pixels and edge pixels, guides the network model to focus more on the boundary regions of camouflaged targets. The camouflaged target detection method based on cascaded frequency domain perception and refined feature guidance enhances the robustness of the model in complex backgrounds, ensuring the integrity and accuracy of the overall localization of camouflaged targets. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the overall method flow of the present invention; Figure 2 This is a schematic diagram of the internal process of the frequency fusion module of the present invention; Figure 3 This is a schematic diagram of the internal process of the aggregation boot decoder of the present invention. Detailed Implementation

[0020] The technical solutions in 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.

[0021] Please see Figure 1 This invention provides a camouflage target detection method based on cascaded frequency domain sensing and refined feature guidance. The method is executed by a computer device or processor. The camouflage target detection network system includes: a feature extraction module, a frequency domain sensing module, a decoding and refinement module, and a loss calculation module.

[0022] The feature extraction module is configured to receive the camouflaged image to be detected. camouflaged images It is a dimension The RGB image data is used by the feature extraction module to extract features from the camouflaged images using a backbone network based on the Spectformer hybrid architecture. Feature encoding is performed, and the backbone network contains convolutional neural network layers and Transformer layers to generate feature maps at four different scales. The feature maps output by the feature extraction module are denoted as follows: ,in, This represents the first-level feature map, which has the highest spatial resolution and contains detailed information about the image; This represents the second-level feature map; This represents the third-level feature map; This represents the fourth-level feature map, which has the lowest spatial resolution and contains the semantic information of the image. The number of channels and the size of the four feature maps change progressively.

[0023] The frequency domain sensing module is connected to the feature extraction module and configured to receive feature maps. The frequency domain sensing module processes each feature map (in (Taking values ​​1, 2, 3, 4) Perform depth-based wavelet transform on the feature map. Decomposed into low-frequency bands and high frequency band Low-frequency banding Through the low-pass filter weight matrix With feature map Perform convolution operations and add bias terms. The obtained high-frequency band Through the high-pass filter weight matrix With feature map Perform convolution operations and add bias terms. The obtained low-frequency band Characterizing the overall contour and color distribution data of the image, high-frequency banding Data that characterizes the edge and texture details of an image.

[0024] The frequency domain sensing module further processes the decomposed frequency domain data using frequency-doubling convolution units. These units include low-frequency and high-frequency branches, with low-frequency band division. Input is fed into the low-frequency branch and passed through the low-frequency convolution kernel. Perform convolution operations and add biases Generate low-frequency features High frequency band Input is fed into the high-frequency branch and passed through a high-frequency convolution kernel. Perform convolution operations and add biases Generate high-frequency features .

[0025] The frequency domain sensing module internally constructs a cascaded frequency domain perceptron, which consists of four cascaded frequency fusion units configured in a bottom-up order (i.e., from...). arrive Aggregate frequency domain features, in the first In the multi-level processing, the frequency fusion unit receives the low-frequency features of the current level. High-frequency characteristics And from the next higher level (the Frequency fusion characteristics (level) For the first level ( The frequency fusion unit only receives... and The cascaded frequency domain perceptron outputs four levels of frequency fusion features. .

[0026] The specific execution steps of the frequency fusion unit are as follows: First, frequency fusion features from the previous level. Perform convolution and normalization to generate intermediate features. .

[0027] Second, intermediate features These are respectively compared with the high-frequency features of the current layer after convolution and normalization. and low-frequency characteristics The data is concatenated along the channel dimension to generate high-frequency input data. and low-frequency input data .

[0028] Third, high-frequency input data is processed using a high-frequency enhancement unit. The high-frequency enhancement unit preserves features through residual blocks and uses a collaborative attention mechanism to calculate spatial and channel weights to generate enhanced high-frequency features. .

[0029] Fourth, use a low-frequency enhancement unit to process low-frequency input data. The low-frequency enhancement unit sequentially performs residual processing and position normalization processing on the data under instance normalization constraints, and then uses a collaborative attention mechanism to calculate weights to generate enhanced low-frequency features. .

[0030] Fifth, it will enhance high-frequency characteristics. intermediate features And enhance low-frequency characteristics Element-wise addition is performed, and the results are fused through convolutional layers to obtain the frequency fusion features of the current layer. .

[0031] The decoding and refinement module is connected to the frequency domain sensing module and configured to receive frequency fusion features. The decoding and refinement module contains four aggregated guided decoding units, configured in a top-down order (i.e., from...). arrive Generate a prediction map.

[0032] In the During the level decoding process, the aggregation-guided decoding unit receives the frequency fusion features of the current level. and from the next level (i.e., deeper levels) Decoding features of ) For the deepest level ( The aggregation-guided decoding unit only receives... The decoding and refinement module ultimately outputs four levels of prediction maps. .

[0033] The specific execution steps of the aggregation boot decoding unit are as follows: First, decoding features from deep layers. Perform an upsampling operation to make its size match the current level features. Consistent.

[0034] Second, the upsampled decoding features With current level features splicing along the channel dimension, and through... The convolutional layer adjusts the number of channels to generate initial fused features. .

[0035] Third, the initial fusion features The input is fed into a variable kernel convolutional (AKConv) layer. The variable kernel convolutional layer dynamically adjusts the sampling position of the convolutional kernel according to the feature content, and outputs the first branch features. Second branch features .

[0036] Fourth, use a partial convolution strategy to process the features of the first branch. Second branch features Specifically, the feature channel is divided into two parts. Convolution is performed on only one part, while the other part remains unchanged. The two parts are then concatenated and fused using pointwise convolution to obtain the first refined feature. Second refinement features Fifth, refine the first feature. Second refinement features Element-wise addition is performed, followed by upsampling, to generate the decoding features for the current level. .

[0037] The loss calculation module is connected to the decoding and refinement module and configured to receive the decoded features of each level. (in (Values ​​range from 1 to 4), because The loss calculation module represents the camouflage target prediction results at this level and compares them with the actual camouflage target label map. For comparison, the loss calculation module uses a hybrid loss function to calculate the total loss value. Total loss value Calculated using the following formula: ; in, This represents the weighted binary cross-entropy loss function, used to calculate the prediction map and the label map. Differences in pixel classification; This represents the weighted intersection-over-union loss function, used to calculate the difference between the predicted map and the label map. Differences in the degree of overlap of the target regions.

[0038] The network minimizes the total loss value using the backpropagation algorithm. This is to update the learnable parameters in the feature extraction module, frequency domain sensing module, and decoding and refinement module.

[0039] See attached document Figure 1 The computer device or processor first acquires the camouflage image to be detected, which is defined as a three-dimensional tensor. ,in This indicates the vertical height in pixels of the camouflaged image. This indicates the horizontal width of the camouflaged image in pixels. Representing the red, green, and blue color channels, in order to adapt to the input requirements of the feature extraction module, the computer device performs size standardization processing on the camouflage image, and uses a bilinear interpolation algorithm to adjust the camouflage image to a preset resolution size. In addition, the computer device performs normalization processing on the camouflage image, mapping the pixel values ​​to a standard normal distribution to eliminate the influence of light intensity differences on the detection results.

[0040] The preprocessed camouflaged image is input into the feature extraction module, which is built on the Spectformer hybrid architecture. The feature extraction module contains four concatenated feature encoding stages, each corresponding to a different feature scale. The design of the feature extraction module adopts a strategy that combines convolutional neural networks and Transformer architecture, taking into account both the efficiency of extracting local detail features and the modeling ability of global semantic features.

[0041] In the first two stages of the feature extraction module, namely the first encoding stage and the second encoding stage, the feature extraction module mainly adopts convolution operation or shallow spectral transformation attention mechanism. Since the target in the camouflaged image has a texture similar to the background and the boundary is blurred, the feature extraction module maintains a high spatial resolution in the shallow network and focuses on capturing texture abrupt changes, edge gradients and small geometric structure information in the camouflaged image.

[0042] In the last two stages of the feature extraction module, namely the third and fourth encoding stages, the feature extraction module adopts the Transformer self-attention mechanism. Since the camouflaged target occupies a large area of ​​the camouflaged image or has separate parts, the feature extraction module uses the self-attention mechanism in the deep network to calculate the long-distance dependencies between pixels, capture the global contextual information and semantic category information of the camouflaged image, and thus locate the overall region of the camouflaged target.

[0043] The feature extraction module outputs four levels of feature maps sequentially through four encoding stages, which are defined as the first-level features. Second-level features Third-level features and fourth-level features First-level features The first-level features are output from the first encoding stage of the feature extraction module. The spatial resolution is one-quarter of the resolution of the input camouflage image, that is... First-level features It contains the richest spatial detail and underlying texture features, but its semantic abstraction is the lowest; second-level features The second-level features are output from the second encoding stage of the feature extraction module. The second-level features are obtained by downsampling and further encoding the output of the first encoding stage. The spatial resolution is one-eighth of the resolution of the input camouflage image, that is... Third-level features Output from the third encoding stage of the feature extraction module. Third-level features. The spatial resolution is one-sixteenth of the resolution of the input camouflage image, that is... Fourth-level features The fourth-level features are output from the fourth encoding stage of the feature extraction module. The spatial resolution is one thirty-second of the input camouflage image resolution, that is... Fourth-level features After multiple downsampling and deep abstraction, it has the largest receptive field and the highest level of semantic information, which can indicate the approximate location of the camouflaged target in the camouflaged image, but it loses some high-frequency boundary details.

[0044] The feature extraction module will generate the first-level features Second-level features Third-level features and fourth-level features As output, it is transmitted to the cascaded frequency domain sensing module for use in subsequent frequency domain decomposition steps.

[0045] See attached document Figure 1 and Figure 2 The cascaded frequency domain sensing module receives first-level features from the feature extraction module. Second-level features Third-level features and fourth-level features The cascaded frequency domain sensing module independently performs frequency domain feature separation and processing operations for each feature level. The following description uses the first... Level features Taking the processing procedure as an example, in which The value can be 1, 2, 3 or 4.

[0046] The cascaded frequency domain sensing module first utilizes a depth-based wavelet transform unit to process the first... Level features The deep wavelet transform unit (DTU) is designed to simulate the human visual system's differentiated perception mechanism of different spatial frequency components. The DTU uses a learnable convolutional filter to decompose the wavelet transform... Level features Separated into the first dimension along the channel dimension Low-frequency characteristic banding and the High-frequency characteristic banding .

[0047] No. Low-frequency characteristic banding Through the low-pass filter weight matrix For the first Level features Perform convolution operations and add low-pass filter bias terms. Generated. (Number) Low-frequency characteristic banding It primarily preserves information about areas with gradual changes in the camouflaged image, including overall color distribution, lighting and shadows, and the macroscopic structural outline of the camouflaged target and background. High-frequency characteristic banding Through the high-pass filter weight matrix For the first Level features Perform convolution operations and add the high-pass filter bias term. Generated. (Number) High-frequency characteristic banding The main focus was on extracting information from drastically changing regions in the camouflaged image, including object edge boundaries, surface texture details, and subtle texture differences between the camouflaged target and the background. The specific calculation formula for depth-based wavelet transform is as follows: ; ; in, This represents the convolution operator.

[0048] Low-pass filter weight matrix and high-pass filter weight matrix The parameters are automatically optimized and updated during network training using the backpropagation algorithm, enabling the frequency domain separation process to adaptively adjust for the camouflage target detection task.

[0049] In obtaining the first Low-frequency characteristic banding and the High-frequency characteristic banding Subsequently, the cascaded frequency domain sensing module uses frequency doubling convolution units to process these two bands in parallel. The frequency doubling convolution unit contains physically separate low-frequency processing branches and high-frequency processing branches, aiming to reduce redundant interference between high-frequency and low-frequency information, and to allocate different computing resources according to the characteristics of different frequency components.

[0050] Frequency doubling convolution unit will Low-frequency characteristic banding Input is fed into the low-frequency processing branch. The low-frequency processing branch utilizes low-frequency convolution kernels. For the first Low-frequency characteristic banding Perform feature mapping and superimpose a low-frequency bias term. , generate the first Low-frequency processing characteristics , No. Low-frequency processing characteristics It further aggregates global semantic and structural information.

[0051] Frequency doubling convolution unit will High-frequency characteristic banding The input is fed into the high-frequency processing branch, which utilizes high-frequency convolution kernels. For the first High-frequency characteristic banding Perform feature mapping and superimpose high-frequency bias terms. , generate the first High-frequency processing characteristics . No. High-frequency processing characteristics It further enhances the expression of edge and texture details.

[0052] ; ; Cascaded frequency domain sensing modules for first-level features Second-level features Third-level features and fourth-level features All samples undergo depth-based wavelet transform and frequency-doubling convolution, ultimately outputting four sets of processed frequency domain feature pairs: , , and These four sets of frequency domain feature pairs will serve as input data for subsequent cascaded frequency fusion steps.

[0053] See attached document Figure 1 The cascaded frequency domain sensing module utilizes four frequency fusion units to construct a cascaded frequency domain perceptron. The cascaded frequency domain sensing module is configured to receive four sets of generated processed frequency domain feature pairs: and , and , and ,as well as and .

[0054] The cascaded frequency domain sensing module adopts a bottom-up cascaded interaction strategy, meaning that the direction of information flow is from the shallow high-resolution feature layer (first level, ...). Towards deeper, low-resolution feature layers (Level 4, This method aims to gradually inject the rich texture and edge details contained in the shallow layer into the semantic features of the deep layer, thereby enhancing the network's ability to perceive the subtle features of the disguised target.

[0055] Cascaded frequency domain sensing modules generate the first Frequency fusion characteristics The process is divided into an initial stage and a cascading stage, among which... The value can be 1, 2, 3 or 4.

[0056] In the initial stage, that is, when At that time, the cascaded frequency domain sensing module uses the first-stage frequency fusion unit to process the first-stage low-frequency processing features. and first-level high-frequency processing features Since the first-level feature is located at the beginning of the cascaded path, the first-level frequency fusion unit only receives the input of the current level, and outputs the first-level frequency fusion feature. First-level frequency fusion features The highest resolution frequency domain synthesis information of the camouflaged image was preserved.

[0057] In the cascading stage, that is when At that time, the cascaded frequency domain sensing module utilizes the first The first-level frequency fusion unit performs cross-level information fusion. The frequency fusion unit at the current level simultaneously receives three inputs: the first input of the current level... Low-frequency processing characteristics The current level High-frequency processing characteristics And from the next higher level (i.e., the first (Level) Frequency fusion characteristics .

[0058] Specifically, the second-level frequency fusion unit receives... , and first-level frequency fusion features Generate second-level frequency fusion features The third-level frequency fusion unit receives , and second-level frequency fusion features Generate third-level frequency fusion features The fourth-level frequency fusion unit receives , and third-level frequency fusion features Generate fourth-level frequency fusion features .

[0059] Through the above cascading interactions, the first Frequency fusion characteristics As prior information, it is passed to the first... Level, making the first Frequency fusion characteristics It not only includes frequency domain information at the current scale, but also aggregates detailed information from shallower levels. The cascaded computation process of the cascaded frequency domain sensing module is defined by the following formula: ; in, This indicates the fusion operation performed by the frequency fusion unit.

[0060] The cascaded frequency domain sensing module ultimately outputs four levels of frequency fusion features. This set of features will be transmitted to the subsequent aggregation-guided decoding module.

[0061] See attached document Figure 2 The frequency fusion module is configured to enhance and fuse the input frequency domain features. The frequency fusion module contains three parallel data processing paths: a cross-level feature processing path, a high-frequency enhancement processing path, and a low-frequency enhancement processing path. The following description uses the first path as an example. Taking the processing of the frequency fusion module as an example, in which... The value can be 2, 3 or 4 (when At this time, cross-level feature processing paths are not executed; only the features at the current level are processed.

[0062] First, the frequency fusion module processes the first feature from the previous level (i.e., the shallower layer with higher resolution) through a cross-level feature processing path. Frequency fusion characteristics ,because Frequency fusion characteristics The spatial resolution is greater than the current number The spatial resolution of the level features is achieved by using a frequency fusion module that combines convolutional or pooling operations with a stride of 2. Convolutional layer, for the first Frequency fusion characteristics Downsampling and channel adjustment are performed to ensure the spatial dimensions match the current layer. Subsequently, feature standardization is achieved through a batch normalization layer, and a nonlinear transformation is introduced using the GeLU activation function. After these processes, the [number]th [level] is generated. Intermediate features across levels The calculation formula for the processing is as follows: ; in, Indicates step size is The downsampling convolution operation, This indicates a batch normalization operation. This indicates the GeLU activation function operation.

[0063] Next, the frequency fusion module constructs the input data for high-frequency and low-frequency enhancement. The frequency fusion module then performs the following steps on the current level: High-frequency processing characteristics and the Low-frequency processing characteristics Perform convolution, batch normalization, and GeLU activation operations respectively.

[0064] The frequency fusion module will process the first... High-frequency processing characteristics With the Intermediate features across levels The data is stitched together along the channel dimension to generate high-frequency enhanced input data. .

[0065] The frequency fusion module will process the first... Low-frequency processing characteristics With the Intermediate features across levels The data is stitched together along the channel dimension to generate low-frequency enhanced input data. The relevant calculation formulas are as follows: ; ; in, This indicates a channel splicing operation.

[0066] Subsequently, the frequency fusion module uses the high-frequency enhancement unit to process the high-frequency enhancement input data. The high-frequency enhancement unit is designed to preserve and enhance the texture and edge details of camouflaged targets. The high-frequency enhancement unit first enhances the input data... The input is fed into the residual block, where a skip connection structure prevents gradient vanishing and preserves the original information of high-frequency features. Next, the high-frequency enhancement unit inputs the output of the residual block into the collaborative attention module. The collaborative attention module simultaneously calculates attention weights in both the channel and spatial dimensions, weighted and filtering out the high-frequency feature responses. The high-frequency enhancement unit outputs enhanced high-frequency features. The calculation formula for high-frequency enhancement processing is as follows: ; Meanwhile, the frequency fusion module uses the low-frequency enhancement unit to process the low-frequency enhancement input data. The low-frequency enhancement unit is designed to optimize global structural information and suppress background noise interference. Since the low-frequency statistical characteristics of the camouflaged target and the background are highly similar, the low-frequency enhancement unit introduces a specific normalization strategy.

[0067] The low-frequency enhancement unit first inputs low-frequency enhancement data. The input is fed into the instance normalized residual block, which normalizes the image within each sample instance and independently adjusts the contrast information.

[0068] Subsequently, the low-frequency enhancement unit performs positional normalization on the features. Positional normalization standardizes the statistical distribution of the channel dimension in spatial location to smooth low-frequency fluctuations in the background region. The low-frequency enhancement unit then inputs the output of the positional normalization to the collaborative attention module, focusing on the low-frequency regions that constitute the target subject. The low-frequency enhancement unit outputs enhanced low-frequency features. The calculation formula for low-frequency enhancement processing is as follows: ; Finally, the frequency fusion module performs feature fusion operations, which enhances the high-frequency features. Enhance low-frequency characteristics and the Intermediate features across levels Perform element-wise addition, and the result is obtained by... The convolutional blocks, composed of convolution, batch normalization, and GeLU activation functions, are integrated to output the final [number of] [units]. Frequency fusion characteristics The formula for feature fusion is as follows: ; in, This indicates an element-wise addition operation.

[0069] No. Frequency fusion characteristics It is then transmitted to the next-level frequency fusion module and the corresponding aggregation guidance decoding module.

[0070] See attached document Figure 3 The aggregation-guided decoding module receives the first-level frequency fusion features from the cascaded frequency domain sensing module. Second-level frequency fusion characteristics Third-level frequency fusion characteristics and fourth-level frequency fusion features The aggregation-guided decoding module contains four aggregation-guided decoding units, each corresponding to one of the four feature levels. The module employs a top-down decoding strategy, progressing from deep networks containing high-level semantic information to shallower networks containing richer detailed information, sequentially generating the fourth-level predicted feature map. Third-level predicted feature map Second-level predicted feature map and first-level predicted feature map .

[0071] The aggregated boot decoding module is for the first Level Predictive Feature Map (in The generation process (taking values ​​of 1, 2, 3, or 4) is defined as follows: When At this point, at the deepest level, the aggregation-guided decoding unit directly addresses the fourth-level frequency fusion features. Decoding is performed to generate a fourth-level predicted feature map. ,when At that time, the aggregation-guided decoding unit performs feature guidance and aggregation operations, the first... The level-based aggregation-guided decoding unit receives two inputs: the first input of the current level. Frequency fusion characteristics and from the next level (i.e., deeper levels) ) Level Predictive Feature Map .

[0072] The aggregation guide decoding unit first performs the first Level Predictive Feature Map Perform an upsampling operation, and use the bilinear interpolation algorithm to... Level Predictive Feature Map Spatial resolution magnified to the same level as the first Frequency fusion characteristics Consistent.

[0073] Subsequently, the aggregation-guided decoding unit will upsample the first... Level Predictive Feature Map With the Frequency fusion characteristics The features are concatenated along the channel dimension, and the concatenated features are then processed by a convolutional kernel of size [size missing]. The convolutional layers are used for initial fusion to generate initial fusion features. This step injects deep semantic context information into shallow detailed features, initially fusing features. The calculation formula is as follows: ; in, Indicates an upsampling operation; This indicates a channel splicing operation. express Convolution operation.

[0074] Next, the aggregation-guided decoding unit uses variable kernel convolution to process the initial fused features. Variable kernel convolution can adjust the initial fusion features. The spatial distribution characteristics dynamically adjust the sampling position of the convolution kernel to adapt to the irregular and varied edge shapes of the camouflaged target. Variable kernel convolution will initially fuse features. The mapping is represented by two parallel branch features, denoted as the first branch feature. Second branch features The calculation formula is as follows: ; To reduce the computational complexity of the decoding process and improve inference speed, the aggregation guides the decoding unit to analyze the features of the first branch. Second branch features Apply partial convolution strategies separately for the first branch features The aggregation-guided decoding unit divides its channel into the first sub-parts that participate in the computation. and the second sub-part that remains unchanged The aggregation-guided decoding unit only works on the first sub-part. implement Convolution operation, the result of which is invariant to the second sub-part The data is reassembled, and the resulting data is processed by pointwise convolution. The convolutional, batch normalization, and GeLU activation functions are integrated to generate the first refined feature. For the second branch features The aggregation-guided decoding unit performs the same partial convolution processing flow to generate the second refined features. The calculation formula for partial convolution processing is as follows: ; in, This indicates the partial convolution operation process.

[0075] Finally, the aggregation-guided decoding unit will refine the first feature. Second refinement features Element-wise addition is performed, and the resulting features are upsampled to restore them to the target resolution, generating the [number]th feature. Level Predictive Feature Map The final output calculation formula is as follows: ; in, This indicates element-wise addition.

[0076] The first-level predicted feature map is finally output by the aggregation-guided decoding module. This refers to the camouflaged target detection results output by the network. Additionally, intermediate generated... It is also retained for subsequent calculation of deep supervision loss.

[0077] See attached document Figure 1 The loss calculation module performs error calculation and parameter optimization tasks during network training. The loss calculation module is configured to receive the first-level predicted feature map output from the aggregation-guided decoding module. Second-level predicted feature map Third-level predicted feature map and the fourth-level predicted feature map Simultaneously, the loss calculation module obtains the real label map corresponding to the camouflaged image. Real label image It is a binary image, where areas with a pixel value of 1 represent areas where camouflaged targets exist, and areas with a pixel value of 0 represent background areas.

[0078] The loss calculation module employs a deep supervision strategy, constraining not only the final output first-level predicted feature map. Instead, it simultaneously supervises the output of each intermediate level in the decoding process. Specifically, the loss calculation module will use the first-level predicted feature map... Second-level predicted feature map Third-level predicted feature map and the fourth-level predicted feature map Compared with the real label image Before comparison, the loss calculation module uses downsampling to compare the true label images. Adjust the resolution to match the predicted feature maps at each level, or use upsampling to restore the predicted feature maps at each level to match the true label maps. Same resolution ensures dimension matching.

[0079] The loss calculation module constructs a hybrid loss function, which is a linear combination of a weighted binary cross-entropy loss function and a weighted intersection-over-union (IoU) loss function. This combination aims to balance pixel-level classification accuracy with image-level overall structural integrity. The total loss function... The calculation formula is as follows: ; in, This represents the calculated total loss value; The index represents the feature level, with a value ranging from 1 to 4; Indicates the first Level prediction feature map; Represents the actual label image; This represents the operation of the weighted binary cross-entropy loss function; This indicates the operation of the weighted intersection-union loss function.

[0080] Regarding the weighted binary cross-entropy loss function The loss calculation module uses a weighted binary cross-entropy loss function to calculate the classification error of each pixel in the predicted feature map. Unlike the standard binary cross-entropy loss, the weighted binary cross-entropy loss function introduces a weight factor. The loss calculation module calculates the weight based on the classification difficulty of the pixel. For difficult pixels that are hard to distinguish, such as those located at the edges of camouflaged targets or in areas with blurred textures, larger weight values ​​are assigned; for easy pixels that are easy to distinguish, such as those located in flat background areas, smaller weight values ​​are assigned. In this way, the loss calculation module guides the feature extraction module, the cascaded frequency domain perception module, and the aggregation-guided decoding module to pay more attention to the boundary details of the camouflaged target and reduce edge blurring.

[0081] Regarding the weighted intersection-and-comparison loss function The loss calculation module uses the weighted intersection-union loss function to calculate the overlap error between the predicted region and the real region. The weighted intersection-union loss function focuses on the global structural consistency of the predicted target. The loss calculation module calculates the ratio of the intersection and union of the predicted mask and the real mask, and performs weighted optimization by combining the weight graph, so that the network can better capture the overall shape of the disguised target and prevent the prediction results from having internal holes or structural breaks.

[0082] The total loss value was calculated. Then, the computer equipment uses the backpropagation algorithm to calculate the total loss value. Relative to the gradients of all learnable parameters in the network, the computer device then uses a gradient descent optimizer to update the weight parameters in the feature extraction module, the cascaded frequency domain sensing module, and the aggregation-guided decoding module based on the calculated gradients, thereby reducing the total loss value in the next iteration. Until the network converges.

Claims

1. A camouflaged target detection method based on cascaded frequency domain sensing and refined feature guidance, characterized in that, Includes the following steps: S1. The computer device acquires the disguised image to be detected and inputs the disguised image into the feature extraction module to generate four feature maps at different scales; S2. The frequency domain sensing module is used to perform frequency domain feature separation on the feature maps of the four different scales respectively to generate corresponding low-frequency feature bands and high-frequency feature bands. The frequency doubling convolution strategy is used to process the low-frequency feature bands and the high-frequency feature bands respectively to obtain low-frequency processed features and high-frequency processed features. S3. Using the cascaded frequency domain perceptron constructed in the frequency domain perception module, the low-frequency processing features and the high-frequency processing features are aggregated across levels in a bottom-up order to generate four levels of frequency fusion features. S4. Using the aggregation-guided decoding unit built in the decoding and refinement module, the frequency fusion features are received in a top-down order, and the refined features are combined with the decoded features from the deep layer. During the refinement process, the features are processed using a variable kernel convolution strategy and a partial convolution strategy to generate four levels of predicted feature maps. S5. Output the predicted feature map of the first level as the camouflage target detection result, and use the loss calculation module to calculate the loss value between the predicted feature map of all levels and the real label map, and update the network parameters through the backpropagation algorithm.

2. The camouflaged target detection method based on cascaded frequency domain sensing and refined feature guidance according to claim 1, characterized in that, In step S2, the step of performing frequency domain feature separation on the four feature maps at different scales using the frequency domain sensing module includes: By using the low-pass filter weight matrix in the depth wavelet transform unit to perform convolution operation with the feature map and superimposing the bias term, low-frequency feature bands are generated, preserving the background and target structure information that change gradually in the camouflaged image. By performing convolution operations between the high-pass filter weight matrix in the depth wavelet transform unit and the feature map, and superimposing bias terms, high-frequency feature bands are generated to extract information on drastically changing object boundaries and subtle texture differences in camouflaged images. The low-pass filter weight matrix and the high-pass filter weight matrix are learnable parameters during the network training process.

3. The camouflaged target detection method based on cascaded frequency domain sensing and refined feature guidance according to claim 1, characterized in that, In step S3, the cross-level aggregation of the low-frequency processing features and the high-frequency processing features in a bottom-up order includes: For the first level, the frequency fusion unit receives the low-frequency processing features and the high-frequency processing features of the first level to generate the frequency fusion features of the first level. For the second to fourth levels, the frequency fusion unit is used to receive the low-frequency processing characteristics of the current level, the high-frequency processing characteristics of the current level, and the frequency fusion characteristics from the previous level. The frequency fusion unit uses the frequency fusion features from the previous level as prior information and fuses them with the low-frequency processing features and high-frequency processing features of the current level to generate the frequency fusion features of the current level.

4. The camouflaged target detection method based on cascaded frequency domain sensing and refined feature guidance according to claim 3, characterized in that, The frequency fusion unit uses the frequency fusion features from the previous level as prior information and fuses them with the low-frequency processing features and high-frequency processing features of the current level. Specifically, this includes: The frequency fusion features from the previous level are downsampled and channel adjusted, and then normalized and activated to generate cross-level intermediate features. The cross-level intermediate features and the high-frequency processing features of the current level are concatenated in the channel dimension and input into the constructed high-frequency enhancement unit. The features are preserved through the residual connection structure, and the spatial weights and channel weights are calculated using the collaborative attention mechanism to generate enhanced high-frequency features. The cross-level intermediate features and the low-frequency processing features of the current level are concatenated in the channel dimension and input into the constructed low-frequency enhancement unit. The residual processing of instance normalization constraints, the position normalization processing and the collaborative attention mechanism processing are executed in sequence to generate enhanced low-frequency features. The enhanced high-frequency features, the cross-level intermediate features, and the enhanced low-frequency features are added element-wise and then fused through a convolutional layer to obtain the frequency fusion features of the current level.

5. The camouflaged target detection method based on cascaded frequency domain sensing and refined feature guidance according to claim 1, characterized in that, In step S4, the process of using the aggregation-guided decoding unit constructed in the decoding and refinement module to receive the frequency fusion features in a top-down order and perform guided refinement in combination with the decoding features from deeper layers includes: For the fourth level, the frequency fusion features of the fourth level are directly decoded to generate the predicted feature map of the fourth level. For the first to third levels, the aggregation-guided decoding unit receives the frequency fusion features of the current level and the predicted feature map from the next level. An upsampling operation is performed on the predicted feature map from the next level to make its size consistent with the frequency fusion feature of the current level, and the two are concatenated in the channel dimension and then generated as an initial fusion feature through a convolutional layer.

6. The camouflaged target detection method based on cascaded frequency domain sensing and refined feature guidance according to claim 5, characterized in that, The process of using a variable kernel convolution strategy during the thinning process includes: The initial fused features are input into a variable kernel convolutional layer; The variable kernel convolutional layer dynamically adjusts the sampling position of the convolutional kernel according to the spatial distribution of the initial fusion features, generating sampling coordinates of arbitrary shapes to match the edge morphology of the camouflaged target; The initial fused features are mapped into parallel first-branch features and second-branch features through the variable kernel convolutional layer.

7. The camouflaged target detection method based on cascaded frequency domain sensing and refined feature guidance according to claim 6, characterized in that, The process of using partial convolution strategies to process features during the refinement process includes: Perform partial convolution operations on the first branch features and the second branch features respectively; The feature channel is divided into a first sub-part that participates in the calculation and a second sub-part that remains unchanged; Convolution operation is performed only on the first sub-part, the result is re-concatenated with the second sub-part, and then fused by point-by-point convolution to generate the first and second thinning features respectively; The first refined feature and the second refined feature are added element-wise, and an upsampling operation is performed to generate the predicted feature map of the current level.

8. The camouflaged target detection method based on cascaded frequency domain sensing and refined feature guidance according to claim 1, characterized in that, The calculation of the loss value between the predicted feature map and the ground truth label map at all levels using the loss calculation module includes: Construct a hybrid loss function, which consists of a weighted binary cross-entropy loss function and a weighted intersection-over-union (IoU) loss function; The predicted feature maps from the first to the fourth level are compared with the real label map, wherein the real label map is a binarized image marked with the disguised target region; For each level, the difference between the predicted feature map and the real label map in pixel classification is calculated using the weighted binary cross-entropy loss function, and the difference between the predicted feature map and the real label map in target region overlap is calculated using the weighted intersection-union loss function. The total loss value is obtained by summing the calculation results of all levels.

9. The camouflaged target detection method based on cascaded frequency domain sensing and refined feature guidance according to claim 8, characterized in that, The weighted binary cross-entropy loss function assigns weights based on the classification difficulty of pixels. Pixels located at the edge of a camouflaged target or in a region with blurred texture are assigned a greater weight than pixels located in a region with a flat background.

10. The camouflaged target detection method based on cascaded frequency domain sensing and refined feature guidance according to claim 1, characterized in that, The feature extraction module generates the feature map using a backbone network based on a hybrid architecture, wherein the backbone network includes convolutional neural network layers and Transformer layers. In the stage of generating the first-level feature map and the second-level feature map, the convolutional neural network layer is used to extract local texture and edge geometry information; During the generation of the third-level and fourth-level feature maps, the self-attention mechanism in the Transformer layer is used to capture global contextual information and semantic category information to locate the overall region of the camouflaged target.

Images