Camouflage target detection method and system based on frequency domain enhancement and automatic prompt guidance

Abstract

The application provides a camouflage target detection method and system based on frequency domain enhancement and automatic prompt guidance. The method according to the application comprises: acquiring a to-be-detected image; inputting the to-be-detected image into a pre-constructed camouflage target segmentation model to obtain a detection result of the camouflage target; wherein the camouflage target segmentation model is based on a general image segmentation SAM model, and a frequency domain enhancement adapter module and a wavelet prompt module are introduced. By introducing the frequency domain enhancement adapter, the original representation ability of the SAM model is maintained, and the ability to depict the overall structure and edge information of the camouflage target is enhanced. At the same time, through the automatically generated mask prompt, the model is guided to focus on the camouflage target area in position, reducing the interference of the background and non-camouflage targets, so that more stable and accurate camouflage target detection and segmentation can be realized in complex and diverse camouflage scenes, and the application has good practical value.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to a method and system for detecting camouflaged targets based on frequency domain enhancement and automatic prompting guidance. Background Technology

[0002] Camouflaged target detection is an important research area in computer vision, aiming to identify and segment camouflaged targets that are highly similar to the background in color, texture, and structure from an image. Camouflaged targets usually blend into the environment through visual camouflage, making the discriminative information between them and the surrounding background weak, thus posing a high challenge for detection and segmentation in complex scenes.

[0003] Existing camouflaged target detection technologies are mainly based on deep learning models. Early methods typically employed convolutional neural networks, using multi-layer feature extraction and multi-scale feature fusion to enhance the model's ability to perceive targets of different sizes. However, due to the local receptive field characteristics of convolutional operations, deep networks are prone to losing overall structural information of the target during feature extraction, leading to incomplete target boundaries or missegmentation in complex backgrounds.

[0004] To compensate for the limitations of convolutional neural networks in global modeling capabilities, some studies have introduced Transformer-based structures to model long-range dependencies in images through self-attention mechanisms, thereby obtaining richer semantic information. However, these methods still rely on a limited scale of camouflaged target data during training, and their ability to characterize fine-grained edges and structural features remains limited in scenes where the target and background are highly similar.

[0005] In recent years, the general image segmentation model SAM, based on large-scale pre-trained data, has achieved the segmentation of arbitrary targets in images by introducing a cueing mechanism. SAM has strong generalization ability in conventional segmentation scenarios, but because its training data lacks visual corpus information for camouflaged targets, directly applying the SAM model to camouflaged target detection tasks can easily segment non-camouflaged targets or background regions in the image, making it difficult for the model to focus on real camouflaged targets. Existing technologies have proposed two types of improvements from the perspective of efficient parameter tuning: one type guides the model for segmentation through cue engineering. This type of method usually designs a learnable subnetwork to generate cue information related to the camouflaged target from the input image. However, since the parameters of the SAM model itself are frozen, this type of method is highly dependent on external cues. Once the quality of the cues is insufficient, the segmentation performance of the model will drop significantly. The other type of method introduces a lightweight adaptation module into the original model structure to fine-tune the feature representation without updating the backbone parameters. However, most existing adaptation modules only model features in the spatial domain. Due to the sensitivity of spatial domain features to changes in pixel intensity, it is difficult to fully characterize the discriminative features of the camouflaged target at the level of overall structure and edge details, resulting in incomplete target contour representation.

[0006] In summary, existing camouflage target detection methods struggle to simultaneously ensure both the accuracy of target localization and the completeness of boundary representation under complex background conditions. They are unable to effectively capture the structural and edge discrimination information of camouflage targets, resulting in incomplete camouflage target segmentation results and insufficient overall accuracy. Summary of the Invention

[0007] The purpose of this invention is to provide a method and system for detecting camouflaged targets based on frequency domain enhancement and automatic prompting guidance, in order to solve the above-mentioned problems in the prior art.

[0008] This invention provides a method for detecting camouflaged targets based on frequency domain enhancement and automatic prompting guidance, comprising: Acquire the image to be detected; The image to be detected is input into a pre-constructed camouflage target segmentation model to obtain the detection result of the camouflage target; wherein, the camouflage target segmentation model is based on the general image segmentation SAM model and introduces a frequency domain enhancement adaptation module and a wavelet cueing module.

[0009] This invention provides a camouflaged target detection system based on frequency domain enhancement and automatic prompting guidance, comprising: The data module is used to acquire the image to be detected; The detection module is used to input the image to be detected into a pre-constructed camouflage target segmentation model to obtain the detection result of the camouflage target; The camouflaged target segmentation model is based on the general image segmentation SAM model and incorporates a frequency domain enhancement adaptation module and a wavelet cueing module. The SAM model includes an image encoder, a cue encoder, and a mask decoder; The image encoder includes multiple cascaded ViT blocks and multiple frequency domain enhancement adapter modules; each frequency domain enhancement adapter module is embedded after the ViT block at fixed intervals, and each frequency domain enhancement adapter module includes a frequency domain enhancement adapter and a frozen ViT block. The frequency domain enhancement adapter includes a dimension reduction projection layer, a fast Fourier transform unit, an amplitude gating unit, an amplitude channel attention unit, an inverse fast Fourier transform reconstruction unit, and a dimension increase projection layer. The wavelet cueing module is connected to the cueing encoder and includes a multi-level learnable wavelet transform unit and a wavelet feature fusion unit.

[0010] This invention also provides an electronic device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is executed by the processor, it implements the steps of the above-described method for detecting camouflaged targets based on frequency domain enhancement and automatic prompting guidance.

[0011] This invention also provides a computer-readable storage medium storing an information transmission implementation program, which, when executed by a processor, implements the steps of the aforementioned camouflaged target detection method based on frequency domain enhancement and automatic prompting guidance.

[0012] The embodiments of the present invention can include the following beneficial effects: By introducing a frequency domain enhancement adapter, the embodiments of the present invention enhance the ability to characterize the overall structure and edge information of camouflaged targets while maintaining the original representation capabilities of the SAM model; at the same time, through automatically generated mask cues, the model is guided to focus on the camouflaged target area in terms of location, reducing interference from the background and non-camouflaged targets, thereby achieving more stable and accurate camouflaged target detection and segmentation in complex and diverse camouflage scenarios, and has good practical value. Attached Figure Description

[0013] To more clearly illustrate the technical solutions in one or more embodiments of this specification or in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this specification. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0014] Figure 1This is a flowchart of a camouflaged target detection method based on frequency domain enhancement and automatic prompting guidance according to an embodiment of the present invention; Figure 2 This is an overall flowchart of the camouflage target detection method according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the adapter block structure according to an embodiment of the present invention (the adapter block consists of a ViT block and a frequency domain enhancement adapter). Figure 4 This is a schematic diagram of the amplitude gating module structure according to an embodiment of the present invention; Figure 5 This is a schematic diagram of the amplitude channel attention module structure according to an embodiment of the present invention; Figure 6 This is a flowchart of the wavelet prompting module according to an embodiment of the present invention; Figure 7 These are wavelet position hint heatmaps and result diagrams from embodiments of the present invention; Figure 8 This is a schematic diagram of a camouflaged target detection system based on frequency domain enhancement and automatic prompting guidance according to an embodiment of the present invention. Detailed Implementation

[0015] To enable those skilled in the art to better understand the technical solutions in one or more embodiments of this specification, the technical solutions in one or more embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this specification, and not all of the embodiments. Based on one or more embodiments of this specification, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of this document.

[0016] Method Implementation Examples According to embodiments of the present invention, a method for detecting camouflaged targets based on frequency domain enhancement and automatic prompting guidance is provided. Figure 1 This is a flowchart of a camouflaged target detection method based on frequency domain enhancement and automatic prompting guidance according to an embodiment of the present invention, as follows: Figure 1 As shown, the camouflaged target detection method based on frequency domain enhancement and automatic prompting guidance according to an embodiment of the present invention specifically includes: Step S101: Obtain the image to be detected; Step S102: Input the image to be detected into the pre-constructed camouflage target segmentation model to obtain the detection result of the camouflage target; wherein, the camouflage target segmentation model is based on the general image segmentation SAM model and introduces a frequency domain enhancement adaptation module and a wavelet cueing module; The SAM model includes an image encoder, a cue encoder, and a mask decoder; The image encoder includes multiple cascaded ViT blocks and multiple frequency domain enhancement adaptation modules; each frequency domain enhancement adaptation module is embedded after the ViT block at fixed intervals, and each frequency domain enhancement adaptation module includes a frequency domain enhancement adapter and a frozen ViT block; wherein, the frequency domain enhancement adapter includes a dimension reduction projection layer, a fast Fourier transform unit, an amplitude gating unit, an amplitude channel attention unit, an inverse fast Fourier transform reconstruction unit, and a dimension increase projection layer; The wavelet prompting module is connected to the prompting encoder and includes a multi-level learnable wavelet transform unit and a wavelet feature fusion unit; Specifically, inputting the image to be detected into a pre-constructed camouflage target segmentation model to obtain the detection result of the camouflage target includes: The image to be detected is input into the SAM model, and the spatial domain features of the image to be detected are extracted by the image encoder. After normalization, the spatial domain features are transmitted to the frequency domain enhancement adapter for frequency domain transformation and enhancement to obtain the fused spatial-frequency domain feature representation; specifically including: The spatial domain features are normalized and then transmitted to the frequency domain enhancement adapter. The input features are mapped to a low-dimensional embedding space through the dimensionality reduction projection layer. The low-dimensional embedding features are then subjected to a Fast Fourier Transform (FFT) by the Fast Fourier Transform unit and mapped to the frequency domain, decomposing them into amplitude and phase components. The amplitude components are selectively enhanced using the amplitude gating unit. The gated and enhanced amplitude components are then correlated using the amplitude channel attention unit. The enhanced amplitude components with correlation modeling are then synthesized with the phase components. The synthesized frequency domain features are reconstructed back to spatial domain features using the Inverse Fast Fourier Transform (IFFT) reconstruction unit. The reconstructed features are restored to the original feature dimension using the dimensionality increase projection layer and then element-wise added or concatenated with the spatial domain features to obtain the fused spatial-frequency domain feature representation. The phase component remains unchanged during the amplitude component enhancement process; The image to be detected is input into the wavelet cue module, and the image to be detected is recursively decomposed by the multi-level learnable wavelet transform unit to obtain multi-scale frequency domain sub-bands; the wavelet feature fusion unit is used to fuse the frequency domain sub-bands of different scales in a bottom-up manner, and the fused wavelet features are used to generate a position mask cue through convolution mapping, and the position mask cue is input into the cue encoder; The spatial-frequency domain feature representation and the position mask cue input into the cue encoder are transmitted together to the mask decoder, and the segmentation prediction result of the camouflaged target is output as the detection result.

[0017] Furthermore, the method also includes a method for training the camouflaged target segmentation model, specifically including: Obtain the training image and its true segmentation mask; The training image is input into the camouflaged target segmentation model to obtain the predicted segmentation mask; Calculate the joint loss function between the predicted segmentation mask and the real segmentation mask, and optimize the trainable parameters in the camouflaged target segmentation model based on the joint loss function; The expression for the joint loss function is as follows: Formula 1; in, Represents the total loss function. This indicates the weighted average loss. This represents the weighted binary cross-entropy loss. Represents the true segmentation mask of the image. This represents the prediction mask used by the model.

[0018] The following describes in detail the above-mentioned technical solutions of the present invention with reference to the specific circumstances of the camouflaged target detection method based on frequency domain enhancement and automatic prompting guidance in the embodiments of the present invention.

[0019] This invention provides a method for detecting camouflaged targets based on frequency domain enhancement adaptation and automatic cueing guidance. The method is based on the SAM model. The input image is fed into its image encoder to extract semantic features. A frequency domain enhancement adapter is embedded in the image encoder to perform frequency domain transformation on multi-layer spatial domain features and enhance the frequency domain components to strengthen structural and edge features related to the camouflaged target. Simultaneously, a multi-level learnable wavelet transform is performed on the input image to extract multi-scale frequency domain features and generate automatic masking cueing. The automatic masking cueing and the frequency-enhanced semantic features are input together into the mask decoder of the SAM model to output the camouflaged target segmentation result. Specifically, the method includes the following steps: Step 1: Input the image into the image encoder of the SAM model to obtain semantic features; Step 2: Embed multiple frequency domain enhancement adapters in the image encoder to perform frequency domain enhancement processing on multi-layer spatial domain features, and fuse the frequency domain enhanced features with the original spatial features to obtain a deep spatial-frequency domain feature representation; The frequency domain enhancement adapter includes dimension reduction projection, fast Fourier transform, amplitude gating module, amplitude channel attention module, inverse fast Fourier transform reconstruction and dimension increase projection. It performs frequency domain transformation on the spatial domain features extracted by the image encoder and performs gating and channel attention enhancement on the amplitude components to enhance the structural and edge features of the camouflaged target while maintaining the original spatial structure information. The amplitude gating module consists of two convolutional layers, a Gaussian error linear unit (GELU) activation function, and a Sigmoid activation function. By learning gating weights from the input features, the amplitude gating module enables the network to adaptively focus on important amplitude components, preserving important information about the camouflaged target while suppressing irrelevant elements. The amplitude channel attention module feeds the gated and enhanced amplitude features into a 1×1 convolution, and then reshapes them to generate three feature maps. ,in Subsequently, in and Matrix multiplication is used between the transposes, and the softmax function is applied to calculate the amplitude attention map. Then, the feature map Multiply by the attention matrix and transform the output shape to Finally, through learnable scaling parameters Obtain the final output : ; The frequency domain enhancement adapter and the ViT block in SAM together constitute the adapter block, and the adapter blocks are inserted at fixed intervals, specifically one adapter block after every three ViT blocks. Furthermore, to achieve efficient parameter tuning, the parameters of the ViT blocks in the image encoder are kept frozen, and only the parameters of the frequency domain enhancement adapter are updated.

[0020] The frequency domain enhancement adapter is placed after the multi-head attention layer and runs in parallel with the multilayer perceptron structure to achieve interactive fusion of frequency domain information and spatial information.

[0021] Step 3: Input the input image into the wavelet cueing module to obtain multi-level wavelet subbands and adaptively fuse them to generate a position mask cue; Among them, the wavelet prompting module recursively decomposes the input image through multi-level learnable wavelet transform to obtain multi-scale frequency domain information, and fuses the frequency domain components of different scales through a bottom-up wavelet fusion module to highlight the location of the camouflaged target. The multi-level learnable wavelet transform defines high- and low-frequency learnable filters and forms a two-dimensional wavelet kernel through vector outer product, decomposing the image into low-frequency components, horizontal high-frequency components, vertical high-frequency components, and diagonal high-frequency components. The wavelet fusion module fuses the frequency domain components at different levels to form a mask location cue, which guides the SAM model to focus on the segmentation of the camouflaged target region, thereby reducing the interference between the non-camouflaged target and the background.

[0022] Step 4: Input the mask hints generated in Step 3 and the spatial-frequency features obtained in Step 2 into the mask decoder, and output the segmentation result of the camouflaged target. ; Step 5: Optimize and train the model using a joint loss function, which includes a weighted intersection-over-union loss. Weighted binary cross-entropy loss : ; in, This represents the true segmentation mask of the image.

[0023] The embodiments of the present invention will be described in detail below: 1. General Overview of the Method like Figure 2 As shown in the figure, this invention presents a method for detecting camouflaged targets based on frequency domain enhancement adaptation and automatic prompting guidance. This method makes full use of the image representation capabilities of the visual large model (SAM), integrates task-related features and enhances the discriminative differences between camouflaged targets and the background, guides and adapts the SAM, thereby achieving accurate localization and segmentation of camouflaged targets.

[0024] Specifically, the image The image encoder of SAM is used to obtain the semantic feature representation of the camouflaged target, where and These represent the height and width of the input image, respectively. To incorporate the frequency domain information of the camouflaged target into a general image segmentation model, this invention proposes a frequency domain-aware adaptation strategy. A frequency domain enhancement adapter is embedded in the SAM image encoder, enabling it to enhance the frequency domain features of multiple spatial domain layers in the image encoder, thereby achieving frequency domain modeling of the camouflaged target features. Simultaneously, the input image... The input image is fed into the wavelet suggestion module, which performs wavelet decomposition on the input image to obtain multi-scale wavelet sub-band features, and generates automatic location suggestions based on the wavelet sub-band features. This guides the SAM model to focus on segmenting the camouflaged target region. Subsequently, location cues are provided. The image encoder, which serves as the input to the masking cue, also feeds semantic features from the image encoder into the masking decoder to generate the final camouflaged target segmentation prediction. The embodiments of the present invention utilize a joint weighted intersection-exchange ratio loss. Weighted binary cross-entropy loss Optimize the training: ; In the formula, Represents the total loss function. Represents the true segmentation mask of the image. This represents the camouflaged target segmentation result (predicted mask) predicted by the model.

[0025] To accelerate the training process while preserving the general visual knowledge already learned by SAM, this embodiment of the invention freezes the parameters of all ViT modules in the image encoder during model training, and only updates the parameters of the frequency domain enhancement adaptation module, the proposed wavelet cueing module, the cueing encoder, and the mask decoder.

[0026] 2. Frequency Domain Enhancement Adapter This invention introduces a frequency domain enhancement adapter into the image encoder of the SAM model to perform frequency domain modeling and enhancement on the multi-layer features extracted during the encoding process, thereby improving the model's ability to express the structure and edge information of camouflaged targets.

[0027] In the SAM model, the image encoder consists of multiple ViT blocks connected in series, possessing strong image pattern representation and long-distance relationship modeling capabilities. However, due to the lack of specific visual corpora for camouflaged targets, directly utilizing the spatial domain features extracted by the original encoder is insufficient to fully characterize the structural differences between the camouflaged target and the background. Therefore, this embodiment of the invention periodically inserts an adapter module into the SAM image encoder to introduce supplementary prior information for the camouflaged target task while keeping the original model parameters frozen.

[0028] like Figure 3 As shown, the adapter module consists of the Frequency Domain Enhancement Adapter (FEA) designed in this embodiment of the invention and the frozen ViT module. The FEA is connected in parallel with the original MLP, enabling the frequency domain enhanced features to participate in subsequent feature modeling in conjunction with the spatial domain features. In the FEA, the intermediate features normalized from the image encoder are first processed, and then mapped to a low-dimensional embedding space through a dimension reduction projection layer. ,in Represents the original feature dimension. The low-dimensional embedding features are then subjected to a Fast Fourier Transform (FFT) to map them to a frequency domain representation, decomposing them into amplitude and phase components. For the amplitude component, the frequency domain enhancement adapter selectively modulates it using an amplitude-gated module to highlight frequency information relevant to the overall structure and edges of the camouflaged target while suppressing background interference. For example... Figure 4 As shown, the amplitude gating module consists of two... The system consists of a convolutional layer, a Gaussian Error Linear Unit (GELU) activation function, and a Sigmoid activation function. By learning gating weights from the input features, the amplitude gating module enables the network to adaptively focus on important amplitude components, preserving crucial information about the camouflaged target while suppressing irrelevant elements.

[0029] Building upon this, an amplitude channel attention module is introduced to model the correlation of amplitude features between different channels, thereby enhancing the ability to express structural frequency domain features. For example... Figure 5 As shown, the amplitude channel attention module sends the gated and enhanced amplitude features to... In the convolution, the shape is then reshaped to generate three feature maps. ,in , For the number of channels; subsequently, in and Matrix multiplication is used between the transposes, and the softmax function is applied to calculate the amplitude attention map. Then, the feature map Multiply by the attention matrix and transform the output shape to Finally, through learnable scaling parameters Obtain the final output : ; In the formula, The final output feature map combines the original amplitude features and the features enhanced by channel attention. This is an amplitude attention map (amplitude attention matrix), representing the correlation strength (attention weight) between different channels. These are learnable scaling parameters used to control the contribution strength of attention-enhanced features; This represents the original amplitude characteristics, i.e., the characteristics that have been gated and enhanced but have not yet undergone channel attention.

[0030] The phase component remains unchanged during frequency domain enhancement to preserve the original spatial structure information. After frequency domain component enhancement, the enhanced frequency domain features are reconstructed using inverse fast Fourier transform. These enhanced features are then restored to the original feature dimension through an upscaling projection layer. It is then fused with the original spatial domain features and used as input for subsequent ViT blocks.

[0031] In this way, the frequency domain enhancement adapter introduces frequency domain information into the feature extraction process without changing the main structure of the SAM model, thereby achieving collaborative modeling of spatial domain features and frequency domain features, and thus enhancing the model's ability to perceive the global structure and fine-grained edges of camouflaged targets.

[0032] 3. Wavelet Hint Module This invention introduces a wavelet cueing module that automatically generates mask cues to guide the SAM model without requiring manual annotation, providing location information for camouflaged targets and suppressing interference from uncamouflaged areas. The wavelet cueing module includes a multi-level learnable wavelet transform and a wavelet feature fusion module. The multi-level learnable wavelet transform module defines low-frequency and high-frequency learnable wavelet filters. Generate a 2D wavelet kernel through vector outer product: ; in, This is represented as a specific frequency convolution operator. It's a wavelet kernel, which decomposes the image into four components: low frequency... Horizontal high frequency Vertical high frequency and diagonal high frequency The low-frequency and high-frequency filter parameters used in wavelet transform are all learnable parameters, which can adaptively adjust the frequency domain decomposition method according to the data distribution during training, thereby enhancing the ability to express the frequency domain features of camouflaged targets.

[0033] like Figure 6 As shown, a multi-level learnable discrete wavelet transform is performed on the input image, decomposing the input image layer by layer into a low-frequency sub-band and multiple high-frequency sub-bands, where... Indicates the first The image is divided into wavelet subbands. The low-frequency subband is used to characterize the overall structural information of the image, while the high-frequency subband is used to characterize the texture and edge information in the image.

[0034] In the multi-level wavelet decomposition process, the wavelet components of higher levels are recursively generated from the low-frequency components of lower levels. To reduce computational complexity, only the low-frequency components of the highest level are retained for subsequent processing. For wavelet components of the same level, the low-frequency and high-frequency components are enhanced by wavelet feature convolution operations, which include convolution operations and nonlinear activation operations to enhance the discriminative power of features in different frequency bands.

[0035] Subsequently, the wavelet feature fusion unit fuses wavelet features from different levels in a bottom-up manner. Specifically, it concatenates the current layer features with wavelet features from deeper layers and fuses them through a convolutional module to reduce the loss of detail information during cross-layer feature fusion, thereby obtaining a fused feature representation containing multi-scale frequency domain information. Finally, the fused wavelet features are used to generate a mask cue through convolutional mapping, ensuring its size matches the input requirements of the cue encoder in a general image segmentation model. The mask cue is then input into the cue encoder to guide the segmentation decoder to focus on the disguised target region.

[0036] like Figure 7As shown, through the above-mentioned prompting module, the embodiments of the present invention can automatically mine the frequency domain difference between the camouflaged target and the background under complex background conditions, generate stable target position prompts, thereby improving the localization accuracy and segmentation stability of the general image segmentation model in the camouflaged target segmentation task.

[0037] 4. The key points of the embodiments of the present invention are mainly reflected in the following three aspects: A. A Network Framework Based on Frequency Domain Enhancement Adaptation and Automatic Prompt Guidance: A holistic framework for camouflaged target segmentation based on the SAM model was constructed. While freezing the main parameters of the original model, a frequency domain enhancement adapter and an automatic prompting module were introduced to achieve effective adaptation and guidance for the camouflaged target detection task. This holistic framework incorporates frequency domain information at the feature level to enhance the structural expressiveness of camouflaged targets, and guides the model to focus on the camouflaged target region through automatically generated mask prompts, thereby achieving stable camouflaged target segmentation under complex background conditions.

[0038] B. Frequency Domain Enhancement Adapter: Since the differences between the camouflaged target and the background in overall structure and edge information are mainly reflected in frequency domain features, a frequency domain enhancement adapter is embedded in the image encoder of the SAM model to perform frequency domain transformation on multi-layer deep features and selectively enhance the frequency domain amplitude components. Through a joint mechanism of amplitude gating and amplitude channel attention, the discriminative frequency components related to the camouflaged target are enhanced while preserving phase information to maintain the original spatial structure, thereby achieving the synergistic expression of frequency domain features and spatial domain features.

[0039] C. Automatic Hint Module: This module performs frequency domain decomposition on the input image using multi-level learnable wavelet transform, extracting low-frequency structural information as well as high-frequency texture and edge information at different scales. It then fuses these multi-scale wavelet features in a bottom-up approach to generate stable camouflage target location hints. The generated mask hints can be used as input to a general image segmentation model without manual annotation, guiding the model to focus on camouflaged target regions under complex background conditions and reducing interference from uncamouflaged areas.

[0040] In summary, the camouflaged target detection method based on frequency domain enhancement adaptation and automatic prompting proposed in this invention has achieved significant performance improvements on several mainstream camouflaged target detection datasets, such as... Figure 7As shown in the figure, the results demonstrate that the method proposed in this embodiment exhibits superior segmentation performance in various complex camouflage scenarios, including those with varying scales, severe occlusion, or complex topological structures. This is attributed to the frequency domain enhancement adapter designed in this embodiment, which integrates the captured deep frequency domain information into the SAM to compensate for the single representation of spatial domain features, thereby enabling the model to extract subtle discriminative features of these camouflaged targets. Furthermore, the wavelet cueing module proposed in this embodiment effectively reduces the adverse effects of interference regions through different levels of frequency domain components, thereby guiding the model to perceive the location of camouflaged targets.

[0041] System Implementation Examples According to embodiments of the present invention, a camouflaged target detection system based on frequency domain enhancement and automatic prompting guidance is provided. Figure 8 This is a schematic diagram of a camouflaged target detection system based on frequency domain enhancement and automatic prompting guidance according to an embodiment of the present invention, as shown below. Figure 8 As shown, the camouflaged target detection system based on frequency domain enhancement and automatic prompting guidance according to an embodiment of the present invention specifically includes: Data module 80 is used to acquire the image to be detected; The detection module 82 is used to input the image to be detected into a pre-constructed camouflage target segmentation model to obtain the detection result of the camouflage target; The camouflaged target segmentation model is based on the general image segmentation SAM model and incorporates a frequency domain enhancement adaptation module and a wavelet cueing module. The SAM model includes an image encoder, a cue encoder, and a mask decoder; The image encoder includes multiple cascaded ViT blocks and multiple frequency domain enhancement adapter modules; each frequency domain enhancement adapter module is embedded after the ViT block at fixed intervals, and each frequency domain enhancement adapter module includes a frequency domain enhancement adapter and a frozen ViT block. The frequency domain enhancement adapter includes a dimension reduction projection layer, a fast Fourier transform unit, an amplitude gating unit, an amplitude channel attention unit, an inverse fast Fourier transform reconstruction unit, and a dimension increase projection layer. The wavelet prompting module is connected to the prompting encoder and includes a multi-level learnable wavelet transform unit and a wavelet feature fusion unit; Specifically, the detection module is used for: The image to be detected is input into the SAM model, and the spatial domain features of the image to be detected are extracted by the image encoder. After normalization, the spatial domain features are transmitted to the frequency domain enhancement adapter for frequency domain transformation and enhancement to obtain the fused spatial-frequency domain feature representation; specifically including: The spatial domain features are normalized and then transmitted to the frequency domain enhancement adapter. The input features are mapped to a low-dimensional embedding space through the dimensionality reduction projection layer. The low-dimensional embedding features are then subjected to a Fast Fourier Transform (FFT) by the Fast Fourier Transform unit and mapped to the frequency domain, decomposing them into amplitude and phase components. The amplitude components are selectively enhanced using the amplitude gating unit. The gated and enhanced amplitude components are then correlated using the amplitude channel attention unit. The enhanced amplitude components with correlation modeling are then synthesized with the phase components. The synthesized frequency domain features are reconstructed back to spatial domain features using the Inverse Fast Fourier Transform (IFFT) reconstruction unit. The reconstructed features are restored to the original feature dimension using the dimensionality increase projection layer and then element-wise added or concatenated with the spatial domain features to obtain the fused spatial-frequency domain feature representation. The phase component remains unchanged during the amplitude component enhancement process; The image to be detected is input into the wavelet cue module, and the image to be detected is recursively decomposed by the multi-level learnable wavelet transform unit to obtain multi-scale frequency domain sub-bands; the wavelet feature fusion unit is used to fuse the frequency domain sub-bands of different scales in a bottom-up manner, and the fused wavelet features are used to generate a position mask cue through convolution mapping, and the position mask cue is input into the cue encoder; The spatial-frequency domain feature representation and the position mask cue input into the cue encoder are transmitted together to the mask decoder, and the segmentation prediction result of the camouflaged target is output as the detection result.

[0042] The embodiments of the present invention are system embodiments corresponding to the above method embodiments. The specific operation of each module can be understood by referring to the description of the method embodiments, and will not be repeated here.

[0043] In summary, this invention proposes a camouflaged target detection method based on frequency domain enhancement adaptation and automatic prompting guidance. This method is based on the general image segmentation model SAM. While freezing its backbone parameters, a frequency domain enhancement adapter is introduced into its image encoder to supplement the extracted deep spatial features with frequency domain representation and enhance the structural and edge information of the camouflaged target. Simultaneously, a multi-level learnable wavelet transform is used to automatically generate location prompts for the camouflaged target, guiding the model to focus on the segmentation of the camouflaged target region. This invention requires no manual prompting and can achieve stable and accurate target detection and segmentation in complex and diverse camouflage scenarios, improving segmentation accuracy and boundary integrity. It is suitable for camouflage target detection in various complex scenarios and related applications.

[0044] Device Example 1 This invention provides an electronic device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is executed by the processor, it performs the steps described in the method embodiment.

[0045] Device Example 2 This invention provides a computer-readable storage medium storing an information transmission implementation program, which, when executed by a processor, performs the steps described in the method embodiment.

[0046] The computer-readable storage media described in this embodiment include, but are not limited to, ROM, RAM, disk, or optical disk.

[0047] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for detecting camouflaged targets based on frequency domain enhancement and automatic prompting guidance, characterized in that, include: Acquire the image to be detected; The image to be detected is input into a pre-constructed camouflage target segmentation model to obtain the detection result of the camouflage target; wherein, the camouflage target segmentation model is based on the general image segmentation SAM model and introduces a frequency domain enhancement adaptation module and a wavelet cueing module.

2. The method according to claim 1, characterized in that, The SAM model includes an image encoder, a cue encoder, and a mask decoder; The image encoder includes multiple cascaded ViT blocks and multiple frequency domain enhancement adapter modules; each frequency domain enhancement adapter module is embedded after the ViT block at fixed intervals, and each frequency domain enhancement adapter module includes a frequency domain enhancement adapter and a frozen ViT block. The frequency domain enhancement adapter includes a dimension reduction projection layer, a fast Fourier transform unit, an amplitude gating unit, an amplitude channel attention unit, an inverse fast Fourier transform reconstruction unit, and a dimension increase projection layer. The wavelet cueing module is connected to the cueing encoder and includes a multi-level learnable wavelet transform unit and a wavelet feature fusion unit.

3. The method according to claim 2, characterized in that, The image to be detected is input into a pre-built camouflage target segmentation model to obtain the detection results of the camouflage target, specifically including: The image to be detected is input into the SAM model, and the spatial domain features of the image to be detected are extracted by the image encoder. After the spatial domain features are normalized, they are transmitted to the frequency domain enhancement adapter for frequency domain transformation and enhancement to obtain the fused spatial-frequency domain feature representation. The image to be detected is input into the wavelet cue module, and the image to be detected is recursively decomposed by the multi-level learnable wavelet transform unit to obtain multi-scale frequency domain sub-bands; the wavelet feature fusion unit is used to fuse the frequency domain sub-bands of different scales in a bottom-up manner, and the fused wavelet features are used to generate a position mask cue through convolution mapping, and the position mask cue is input into the cue encoder; The spatial-frequency domain feature representation and the position mask cue input into the cue encoder are transmitted together to the mask decoder, and the segmentation prediction result of the camouflaged target is output as the detection result.

4. The method according to claim 3, characterized in that, After normalizing the spatial domain features, they are transmitted to the frequency domain enhancement adapter for frequency domain transformation and enhancement, resulting in a fused spatial-frequency domain feature representation, specifically including: The spatial domain features are normalized and then transmitted to the frequency domain enhancement adapter. The input features are mapped to a low-dimensional embedding space through the dimensionality reduction projection layer. The low-dimensional embedding features are then subjected to a Fast Fourier Transform (FFT) by the Fast Fourier Transform unit and mapped to the frequency domain, decomposing them into amplitude and phase components. The amplitude components are selectively enhanced using the amplitude gating unit. The gated and enhanced amplitude components are then correlated using the amplitude channel attention unit. The enhanced amplitude components with correlation modeling are then synthesized with the phase components. The synthesized frequency domain features are reconstructed back to spatial domain features using the Inverse Fast Fourier Transform (IFFT) reconstruction unit. The reconstructed features are restored to the original feature dimension using the dimensionality increase projection layer and then element-wise added or concatenated with the spatial domain features to obtain the fused spatial-frequency domain feature representation. The phase component remains unchanged during the amplitude component enhancement process.

5. A method for training a camouflaged target segmentation model according to any one of claims 1 to 4, characterized in that, include: Obtain the training image and its true segmentation mask; The training image is input into the camouflaged target segmentation model to obtain the predicted segmentation mask; Calculate the joint loss function between the predicted segmentation mask and the real segmentation mask, and optimize the trainable parameters in the camouflaged target segmentation model based on the joint loss function; The expression for the joint loss function is as follows: Official 1; in, Represents the total loss function. This indicates the weighted average loss. This represents the weighted binary cross-entropy loss. Represents the true segmentation mask of the image. This represents the prediction mask used by the model.

6. A camouflaged target detection system based on frequency domain enhancement and automatic prompting guidance, characterized in that, include: The data module is used to acquire the image to be detected; The detection module is used to input the image to be detected into a pre-constructed camouflage target segmentation model to obtain the detection result of the camouflage target; The camouflaged target segmentation model is based on the general image segmentation SAM model and incorporates a frequency domain enhancement adaptation module and a wavelet cueing module. The SAM model includes an image encoder, a cue encoder, and a mask decoder; The image encoder includes multiple cascaded ViT blocks and multiple frequency domain enhancement adapter modules; each frequency domain enhancement adapter module is embedded after the ViT block at fixed intervals, and each frequency domain enhancement adapter module includes a frequency domain enhancement adapter and a frozen ViT block. The frequency domain enhancement adapter includes a dimension reduction projection layer, a fast Fourier transform unit, an amplitude gating unit, an amplitude channel attention unit, an inverse fast Fourier transform reconstruction unit, and a dimension increase projection layer. The wavelet cueing module is connected to the cueing encoder and includes a multi-level learnable wavelet transform unit and a wavelet feature fusion unit.

7. The system according to claim 6, characterized in that, The detection module is specifically used for: The image to be detected is input into the SAM model, and the spatial domain features of the image to be detected are extracted by the image encoder. After the spatial domain features are normalized, they are transmitted to the frequency domain enhancement adapter for frequency domain transformation and enhancement to obtain the fused spatial-frequency domain feature representation. The image to be detected is input into the wavelet cue module, and the image to be detected is recursively decomposed by the multi-level learnable wavelet transform unit to obtain multi-scale frequency domain sub-bands; the wavelet feature fusion unit is used to fuse the frequency domain sub-bands of different scales in a bottom-up manner, and the fused wavelet features are used to generate a position mask cue through convolution mapping, and the position mask cue is input into the cue encoder; The spatial-frequency domain feature representation and the position mask cue input into the cue encoder are transmitted together to the mask decoder, and the segmentation prediction result of the camouflaged target is output as the detection result.

8. The system according to claim 7, characterized in that, The detection module is specifically used for: The spatial domain features are normalized and then transmitted to the frequency domain enhancement adapter. The input features are mapped to a low-dimensional embedding space through the dimensionality reduction projection layer. The low-dimensional embedding features are then subjected to a Fast Fourier Transform (FFT) by the Fast Fourier Transform unit and mapped to the frequency domain, decomposing them into amplitude and phase components. The amplitude components are selectively enhanced using the amplitude gating unit. The gated and enhanced amplitude components are then correlated using the amplitude channel attention unit. The enhanced amplitude components with correlation modeling are then synthesized with the phase components. The synthesized frequency domain features are reconstructed back to spatial domain features using the Inverse Fast Fourier Transform (IFFT) reconstruction unit. The reconstructed features are restored to the original feature dimension using the dimensionality increase projection layer and then element-wise added or concatenated with the spatial domain features to obtain the fused spatial-frequency domain feature representation. The phase component remains unchanged during the amplitude component enhancement process.

9. An electronic device, characterized in that, include: The memory, the processor, and the computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the steps of the camouflaged target detection method based on frequency domain enhancement and automatic prompting guidance as described in any one of claims 1-4.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores an information transmission implementation program, which, when executed by a processor, implements the steps of the camouflaged target detection method based on frequency domain enhancement and automatic prompting guidance as described in any one of claims 1-4.

Images