Semi-supervised camouflaged target detection method based on adaptive data selection and text fusion

Abstract

The present invention discloses a semi-supervised disguised target detection method based on adaptive data selection and text fusion. By utilizing adaptive data enhancement technology and text fusion strategies, model training is performed with only a small amount of manually annotated data required. This guides the model to learn the deep features of the disguised target, achieves accurate segmentation of the disguised target that is extremely similar to the background, and generates high-quality segmentation masks. Specifically, the present invention comprises an adaptive data enhancement module and an adaptive data selection module, which adaptively select valuable samples for annotation through adversarial enhancement and sampling strategies; and a text fusion module, which fully utilizes labeled data by combining disguise knowledge and visual-text interaction. Experimental results show that the method proposed in the present invention surpasses all existing semi-supervised disguised target detection methods, effectively demonstrating its effectiveness.

Description

Technical Field

[0001] The present invention belongs to the field of computer vision technology, specifically the field of semi-supervised disguised target detection, and more particularly refers to a semi-supervised disguised target detection method based on adaptive data selection and text fusion. Background Art

[0002] Camouflaged object detection (COD) is a challenging task. Its core goal is to leverage large amounts of pixel-level annotated data to train models to accurately detect and segment camouflaged objects that closely resemble the background. Compared to traditional salient object detection, camouflaged object detection presents more complex challenges: Camouflaged objects often exhibit characteristics that interfere with detection and localization, such as small size, partial occlusion, and high blending with the background. They can even possess the ability to dynamically change their appearance to enhance their concealment. This high similarity between the object and the background in color, texture, and morphology significantly complicates camouflaged object detection. Since segmenting and annotating camouflaged object images is a challenging task, studying how to efficiently train camouflaged object detection models using minimal data has significant practical value.

[0003] Compared to traditional fully supervised disguised object detection methods, semi-supervised disguised object detection effectively alleviates the reliance on labeled data by training a model to segment disguised objects using a small amount of labeled data and a large amount of unlabeled data. Most existing semi-supervised disguised object detection methods employ a teacher-student network framework, training the model by designing data augmentation strategies or constructing regularized loss terms to mitigate the noise problem of pseudo-labels in the teacher model. However, these methods fail to consider the impact of the quality of labeled data on model training and simply select a certain percentage of data as labeled data through random sampling. Furthermore, these methods lack sufficient information mining of the labeled data, resulting in poor performance.

[0004] The task of directional object detection further refines the scope of object detection. This task focuses on identifying and localizing specific objects in images based on specific cues or descriptions. These methods require not only parsing the visual content of the image but also understanding the descriptions or instructions associated with the object, enabling accurate targeting in complex scenarios. Directional disguised object detection combines research findings from both disguised and directional object detection. Existing image-based directional object detection methods use reference images of the disguised object as directional information to capture a common representation of the target object from a reference image composed of salient objects, thereby identifying the specific disguised object. Additionally, some methods attempt to utilize visual language model question answering (VLM) approaches. By designing a series of question-answering cue word templates, they guide the VLM to locate the disguised object and obtain a relevant description. The performance of these approaches is limited by the VLM's ability to identify the specific disguised object in the image and accurately describe its attributes. If the VLM itself fails to locate the specific disguised object in the image, or if it locates an incorrect number of disguised objects, the accuracy of the model's final prediction will be significantly affected. Therefore, accurate and appropriate descriptions of the target image are crucial to the quality of the trained model.

[0005] In summary, in the existing technology, most existing semi-supervised disguised target detection methods select data as labeled data through random sampling without considering the quality of the selected data; for the small amount of selected labeled data, it is difficult for the model to learn knowledge related to camouflage, and there is a lack of sufficient mining of labeled data information, resulting in poor performance. Summary of the Invention

[0006] The main purpose of the present invention is to provide a semi-supervised disguised target detection method based on adaptive data selection and text fusion to solve the problems existing in the prior art. By utilizing adaptive data enhancement technology and text fusion strategy, the model is trained under the condition of only a small amount of manually annotated data, guiding the model to learn the deep features of the disguised target, achieving accurate segmentation of the disguised target that is extremely similar to the background, and generating high-quality segmentation masks.

[0007] In order to achieve the above object, the solution of the present invention is:

[0008] A semi-supervised disguised target detection method based on adaptive data selection and text fusion includes an adaptive data enhancement module, an adaptive data selection module and a text fusion module, and includes the following steps:

[0009] Step 1: Use the adaptive data selection module to score the value of unlabeled images, select high-value images suitable for model training according to the preset scoring threshold, and manually annotate them to build a labeled dataset containing image-level directional text and real image labels;

[0010] Step 2: First, preprocess the manually annotated image to be detected, and then use the adaptive data enhancement module to obtain the adaptive data enhanced image;

[0011] Step 3: Input the adaptive data augmented image into the teacher-student framework, and use the visual feature extraction backbone network to extract the image deep features; the teacher-student framework includes a teacher model and a student model;

[0012] Step 4: Input the image depth features and image-level directional text into the text fusion module to generate text-visual fusion features, and the decoder generates the teacher model prediction results and the student model prediction results;

[0013] Step 5: Use the real image labels and the student model prediction results to calculate the total loss of the entire network, and then adjust the parameters of the visual feature extraction backbone network, adaptive data enhancement module, text fusion module and decoder through backpropagation.

[0014] The step 1 includes the steps of constructing a disguised target detection dataset RefTextCOD containing image-level directional text:

[0015] Step 1.1.1 Perform directional text annotation on four mainstream camouflaged target detection datasets: CHAMELEON, CAMO, COD10K, and NC4K. Obtain the precise foreground region of the camouflaged target in the input image based on the dataset ground truth map.

[0016] Step 1.1.2: Construct corresponding graphic-cue word pairs according to the preset cue word rules;

[0017] Step 1.1.3 Input the original image, the disguised target foreground area, and the prompt word into the Qwen-VL-Chat model and the GPT4-Vision model respectively to obtain multiple sets of candidate prompt words for each set of images;

[0018] Step 1.1.4: Manually refine the candidate prompt words to obtain accurate directional text.

[0019] Step 1.1.5: Organize the results into the format of image-text pairs for the complete dataset.

[0020] Preferably, the preset prompt word rule is: given an original image, and using the corresponding segmentation mask to extract the camouflaged target as an auxiliary positioning image, the prompt word guides the model to gradually identify the foreground of the camouflaged target and its corresponding background physical characteristics, and guides the VLM to aggregate this information to generate the final annotation.

[0021] In step 1, the adaptive data selection module is used to incrementally expand the labeled data set, specifically including:

[0022] Step 1.2.1 For the first trained model, the adaptive data selection module extracts features in the color, texture, and frequency domains, and uses K-Means clustering to classify the data. The samples close to the cluster center are selected as the initial labeled data, and then proceed to step 2. For the retrained model, the structural similarity index and mean absolute error are calculated for all unlabeled images. Teacher model prediction results after inputting the model Prediction results with student model Data scoring between :

[0023] ;

[0024] in, Represents the index of the unlabeled image; and denote the structural similarity index and mean absolute error, respectively; Refers to no mark, superscript Refers to image, subscript Refers to teacher, subscript Refers to student;

[0025] Step 1.2.2: Standardize the data scores of all unlabeled images, scale the data scores to the numerical range [0, 1] using maximum and minimum normalization, and sort the data according to the normalization results;

[0026] Step 1.2.3: Select data symmetrically starting from the middle of the sorted results, manually label the selected data, and then add it to the labeled dataset.

[0027] In step 2, the original pixel values ​​of the manually annotated image to be detected are scaled to the interval [0, 1] and normalized using the preset channel mean and variance; the image is then scaled to 640×640, and the semantic segmentation mask corresponding to the image is also scaled to 640×640; a trainable color enhancer is used with a trainable geometric transformation enhancer Perform data augmentation on the scaled image.

[0028] In step 3, the teacher model and student model of the teacher-student framework are both composed of a visual feature extraction backbone network, a text fusion module and a decoder, and the student model parameters are subjected to exponential moving average to generate the teacher model parameters; the visual feature extraction backbone network uses the Swin pre-trained model to extract image depth features; the decoder uses a BiRefBlock decoder; the visual feature extraction backbone network, the text fusion module and the decoder all participate in model training.

[0029] The process of inputting the image depth features and the image-level directional text into the text fusion module to generate the text-visual fusion features in step 4 specifically includes the following steps:

[0030] Step 4.1 Use the directional text encoder to encode the image-level directional text to obtain text features ; The directional text encoder uses a pre-trained CLIP text encoder with frozen parameters. This text encoder does not participate in model training and is only used to encode text information; superscript Refers to Text; subscript Refers to Labeled, subscript Indicates the index of labeled data;

[0031] Step 4.2 For image depth features , using the cued attention mechanism to generate cued attention queries :

[0032] ;

[0033] ;

[0034] ;

[0035] Among them, the subscript Represents the first level of multi-level features Features represents the clue query vector; represents a learnable query; represents the clue key vector; represents the clue value vector; 、 and represents the linear projection weight; represents the cue attention map; express The transpose of represents the dimension of the attention head; represents the softmax activation function;

[0036] Step 4.3 Calculate the cosine similarity score between the cued attention query and the codebook :

[0037] ;

[0038] in, represents cosine similarity; Represents an element in the codebook; A codebook representing a set of feature vectors storing category-related information;

[0039] Step 4.4 Based on cosine similarity score Perform weighted aggregation on the codebook to generate clue vector :

[0040] ;

[0041] in, Indicates the calculation of the number of elements in a collection;

[0042] Step 4.5: Clue vector Insert Text Feature , obtain enhanced text features :

[0043] ;

[0044] in, Represents a splicing operation;

[0045] Step 4.6 Enhance text features and image depth features Using cross attention mechanism to obtain text-visual fusion features :

[0046] ;

[0047] in, Indicates cross-attention;

[0048] Step 4.7 Fusion of text-visual features Input decoder to generate teacher model prediction results Prediction results with student model .

[0049] Preferably, in step 5, the total loss of the network The calculation formula is:

[0050] ;

[0051] ;

[0052] ;

[0053] ;

[0054] ;

[0055] ;

[0056] ;

[0057] in, represents the supervision loss, which is used to supervise the prediction results of the student model by the real image label; represents a hyperparameter used to balance supervised loss and unsupervised loss; Represents the unsupervised loss, which is used to supervise the prediction results of the teacher model on the prediction results of the student model; represents the supervised segmentation loss, represents the unsupervised segmentation loss, both of which are composed of binary cross entropy loss, intersection-over-union loss, and structural similarity loss; represents the adaptive enhancer loss; represents the binary cross entropy loss; 、 They represent the prediction results of the teacher model and the student model after the labeled image is input into the model; represents the true annotation of the image; represents the loss of the text fusion module; represents the cue attention map; Indicates that the bilinear interpolation algorithm is used to adjust the resolution; represents the mean square error loss; represents the CLIP pre-trained text encoder; The category word that represents the image.

[0058] After adopting the above technical solution, the present invention has the following technical effects:

[0059] (1) The present invention has an adaptive data enhancement module and an adaptive data selection module, which adaptively selects valuable samples for labeling through adversarial enhancement and sampling strategies;

[0060] (2) This paper proposes a text fusion module that fully utilizes labeled data by combining camouflage knowledge and visual-text interaction;

[0061] (3) This paper proposes a new dataset, called RefTextCOD, which contains image-level text annotations of existing mainstream disguised object detection datasets;

[0062] (4) Simulation experiments were conducted with reference to the existing semi-supervised disguised target detection settings. The experimental results show that the method proposed in this paper surpasses all existing semi-supervised disguised target detection methods, effectively proving its effectiveness. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 It is an overall flow chart of the method proposed in a specific embodiment of the present invention.

[0064] Figure 2 It is a preset prompt word construction rule in a specific embodiment of the present invention.

[0065] Figure 3 This is a specific flow chart for constructing a data set in a specific embodiment of the present invention.

[0066] Figure 4 The following are the experimental results of performance comparison between the specific embodiments of the present invention and mainstream camouflaged target detection methods.

[0067] Figure 5 This is a module validity ablation verification of a specific embodiment of the present invention. DETAILED DESCRIPTION

[0068] In order to further explain the technical solution of the present invention, the present invention is described in detail below through specific embodiments.

[0069] refer to Figure 1 As shown, the present invention discloses a semi-supervised disguised target detection method based on adaptive data selection and text fusion, which includes an adaptive data augmentation module (ADA), an adaptive data selection module (ADS) and a text fusion module (TFM), and includes the following steps:

[0070] Step 1: Use the adaptive data selection module to score the value of unlabeled images, select high-value images suitable for model training according to the preset scoring threshold, and manually annotate them to build a labeled dataset containing image-level directional text and real image labels;

[0071] Step 2: First, preprocess the manually annotated image to be detected, and then use the adaptive data enhancement module to obtain the adaptive data enhanced image;

[0072] Step 3: Input the adaptive data augmented image into the teacher-student framework (including the teacher model and the student model), and the visual feature extraction backbone network extracts the image deep features;

[0073] Step 4: Input the image depth features and image-level directional text into the text fusion module to generate text-visual fusion features, and the decoder generates the teacher model prediction results and the student model prediction results;

[0074] Step 5: Use the real image labels and the student model prediction results to calculate the total loss of the entire network (the complete network including all components), and then adjust the parameters of the visual feature extraction backbone network, adaptive data enhancement module, text fusion module and decoder through backpropagation.

[0075] refer to Figure 2-3 As shown in FIG, the above step 1 includes the construction steps of a disguised target detection dataset RefTextCOD containing image-level directional text:

[0076] Step 1.1.1 Perform directional text annotation on four mainstream camouflaged target detection datasets: CHAMELEON, CAMO, COD10K, and NC4K. Obtain the precise foreground region of the camouflaged target in the input image based on the dataset ground truth map.

[0077] Step 1.1.2 According to the preset prompt word rules (see Figure 2 ), construct the corresponding graphic-prompt word pair; where "{}" means replacing the corresponding position according to the answer result of the previous prompt word;

[0078] Step 1.1.3 Input the original image, the disguised target foreground area, and the prompt word into the Qwen-VL-Chat model and the GPT4-Vision model respectively to obtain multiple sets of candidate prompt words for each set of images;

[0079] Step 1.1.4: Manually refine the candidate prompt words to obtain accurate directional text.

[0080] Step 1.1.5: Organize the results into the format of image-text pairs for the complete dataset.

[0081] Furthermore, the above preset prompt word rules are:

[0082] By giving the original image and using the corresponding segmentation mask to extract the camouflaged target as an auxiliary positioning image, the model is guided by prompt words to gradually identify the foreground of the camouflaged target and its corresponding background physical characteristics, and guide the VLM to aggregate this information to generate the final annotation.

[0083] In step 1 above, the adaptive data selection module is used to incrementally expand the labeled dataset, specifically including:

[0084] Step 1.2.1 For the first trained model, the adaptive data selection module extracts features such as color, texture, and frequency domain, and uses K-Means clustering to classify the data. The samples close to the cluster center are selected as the initial labeled data and proceed to step 2. For the retrained model, the structural similarity index and mean absolute error are calculated for all unlabeled images. Teacher model prediction results after inputting the model Prediction results with student model Data scoring between :

[0085] ;

[0086] in, Represents the index of the unlabeled image; and denote the structural similarity index and mean absolute error, respectively; Refers to no mark, superscript Refers to image (graphics), subscript Refers to teacher, subscript Refers to student;

[0087] Step 1.2.2: Standardize the data scores of all unlabeled images, scale the data scores to the numerical range [0, 1] using maximum and minimum normalization, and sort the data according to the normalization results;

[0088] Step 1.2.3: Select data symmetrically starting from the middle of the sorted results, manually annotate the selected data, and then add it to the annotated dataset. The amount of data selected depends on the need to increase the number of annotated images. For example, if the current annotated data is 1% (41 images), and it needs to be increased to 5% (202 images), then 4% (161 images) of the data is symmetrically selected from the middle. In this embodiment, the annotated data is selected in increments of 1%, 5%, and 10%, respectively.

[0089] In step 2 above, the original pixel values ​​of the manually annotated image to be detected are scaled to the interval [0, 1] and normalized using the preset channel mean and variance (the mean and variance of the graph in the ImageNet dataset are used in the embodiment); the image is then scaled to 640×640, and the semantic segmentation mask corresponding to the image is also scaled to 640×640; a trainable color enhancer is used with a trainable geometric transformation enhancer Perform data augmentation on the scaled image (data augmentation includes but is not limited to brightness, contrast, color perturbation, and affine transformation).

[0090] In the above step 3, the teacher model and student model of the teacher-student framework are both composed of a visual feature extraction backbone network, a text fusion module and a decoder, and the student model parameters are subjected to exponential moving average (EMA) to generate the teacher model parameters; the visual feature extraction backbone network uses the Swin pre-trained model to extract image depth features; the decoder uses the BiRefBlock decoder proposed by Peng Zheng et al. in their work "Bilateral reference for high-resolution dichotomous images segmentation" published in "CAAI Artificial Intelligence Research" in 2024 (PengZheng, Dehong Gao, Deng-Ping Fan, Li Liu, Jorma Laaksonen, Wanli Ouyang, andNicu Sebe. Bilateral reference for high-resolution dichotomous imagesegmentation. CAAI Artificial Intelligence Research, 2024.); the visual feature extraction backbone network, text fusion module and decoder all participate in model training.

[0091] In step 4 above, the process of inputting the image depth features and the image-level directional text into the text fusion module to generate the text-visual fusion features specifically includes the following steps:

[0092] Step 4.1 Use the directional text encoder to encode the image-level directional text to obtain text features ; The directional text encoder uses a pre-trained CLIP text encoder with frozen parameters. This text encoder does not participate in model training and is only used to encode text information; superscript Refers to Text (text); subscript Refers to Labeled (labeled data), subscript Indicates the index of labeled data;

[0093] Step 4.2 For image depth features , using the Clue Attention Mechanism to generate a clue attention query :

[0094] ;

[0095] ;

[0096] ;

[0097] Among them, the subscript Represents the first level of multi-level features Features represents the clue query vector (clue queryvector); represents a learnable query; represents the clue key vector; Represents the clue value vector (clue value vector); 、 and represents the linear projection weight; represents the cue attention map; express The transpose of represents the dimension of the attention head; represents the softmax activation function;

[0098] Step 4.3 Calculate the cosine similarity score between the cued attention query and the codebook :

[0099] ;

[0100] in, represents cosine similarity; Represents an element in the codebook; A codebook representing a set of feature vectors storing category-related information;

[0101] Step 4.4 Based on cosine similarity score Perform weighted aggregation on the codebook to generate clue vector :

[0102] ;

[0103] in, Indicates the calculation of the number of elements in a collection;

[0104] Step 4.5: Clue vector Insert Text Feature , obtain enhanced text features :

[0105] ;

[0106] in, Represents a splicing operation;

[0107] Step 4.6 Enhance text features and image depth features Use the Cross-Attention mechanism to obtain text-visual fusion features :

[0108] ;

[0109] in, Indicates cross-attention;

[0110] Step 4.7 Fusion of text-visual features Input decoder to generate teacher model prediction results Prediction results with student model .

[0111] In step 5 above, the total loss of the network is The calculation formula is:

[0112] ;

[0113] ;

[0114] ;

[0115] ;

[0116] ;

[0117] ;

[0118] ;

[0119] in, represents the supervision loss, which is used to supervise the prediction results of the student model by the real image label; represents a hyperparameter used to balance supervised loss and unsupervised loss; Represents the unsupervised loss, which is used to supervise the prediction results of the teacher model on the prediction results of the student model; represents the supervised segmentation loss, represents the unsupervised segmentation loss, both of which are composed of binary cross entropy loss, intersection-over-union loss, and structural similarity loss; represents the adaptive enhancer loss; represents the binary cross entropy loss; 、 They represent the prediction results of the teacher model and the student model after the labeled image is input into the model; represents the true annotation of the image; represents the loss of the text fusion module; represents the cue attention map; Indicates that the bilinear interpolation algorithm is used to adjust the resolution; represents the mean square error loss; represents the CLIP pre-trained text encoder; The category word that represents the image.

[0120] Through the above scheme, the present invention has an adaptive data enhancement module and an adaptive data selection module, which adaptively select valuable samples for annotation through adversarial enhancement and sampling strategies; at the same time, the present invention proposes a text fusion module, which fully utilizes labeled data by combining camouflage knowledge and visual-text interaction; the present invention proposes a new dataset, called RefTextCOD, which contains image-level text annotations of existing mainstream camouflaged target detection datasets; simulation experiments are carried out with reference to the settings of existing semi-supervised camouflaged target detection. The experimental results show that the method proposed by the present invention surpasses all existing semi-supervised camouflaged target detection methods, effectively proving its effectiveness.

[0121] Experimental process:

[0122] This paper follows the common settings in the field of semi-supervised camouflaged target detection. It uses the CAMO training set and the COD10K training set (excluding images that do not contain camouflaged targets) as the basic training set, and divides the labeled data into 1%, 5%, and 10% proportions. The labeled data division is performed according to the adaptive data enhancement module proposed in this paper. For the selection of the test set, the present invention is tested on four datasets: CHAMELEON (containing a total of 76 test images), CAMO-Test (containing a total of 250 test images), COD10K-Test (containing a total of 2026 test images), and NC4K (containing a total of 4121 test images). During the test, two different settings were used for testing:

[0123] (1) All images use fixed directional text;

[0124] (2) All images use detailed directional text.

[0125] The experimental results are as follows Figure 4As shown in the figure, the experimental results show that the proposed method outperforms the existing semi-supervised disguised target detection model in multiple indicators. Compared with previous methods, it improves the mean absolute error (MAE) by 52.0% and the structural similarity score (S-Measure) by 19.1%, which proves the effectiveness of the model proposed in this paper.

[0126] Figure 5 Functional ablation experiments on relevant modules of the method proposed in the present invention show that the present invention has greatly helped to improve the performance of semi-supervised camouflaged target detection, further proving the effectiveness of the method proposed in the present invention.

[0127] The above embodiments and drawings do not limit the product form and style of the present invention. Any appropriate changes or modifications made by ordinary technicians in the relevant technical field should be deemed to be within the patent scope of the present invention.

Claims

1. A semi-supervised disguised target detection method based on adaptive data selection and text fusion, characterized by It contains an adaptive data enhancement module, an adaptive data selection module and a text fusion module, and includes the following steps: Step 1: Use the adaptive data selection module to score the value of unlabeled images, select high-value images suitable for model training according to the preset scoring threshold, and manually annotate them to build a labeled dataset containing image-level directional text and real image labels; Step 2: First, preprocess the manually annotated image to be detected, and then use the adaptive data enhancement module to obtain the adaptive data enhanced image; Step 3: Input the adaptive data augmented image into the teacher-student framework, and use the visual feature extraction backbone network to extract the image deep features; the teacher-student framework includes a teacher model and a student model; Step 4: Input the image depth features and image-level directional text into the text fusion module to generate text-visual fusion features, and the decoder generates the teacher model prediction results and the student model prediction results; Step 5: Use the real image labels and the student model prediction results to calculate the total loss of the entire network, and then adjust the parameters of the visual feature extraction backbone network, adaptive data enhancement module, text fusion module and decoder through backpropagation; The step 4 specifically includes the following steps: Step 4.1 Use the directional text encoder to encode the image-level directional text to obtain text features ; The directional text encoder uses a pre-trained CLIP text encoder with frozen parameters. This text encoder does not participate in model training and is only used to encode text information; superscript Refers to Text; subscript Refers to Labeled, subscript Indicates the index of labeled data; Step 4.2 For image depth features , using the cued attention mechanism to generate cued attention queries : ； ； ； Among them, the subscript Represents the first level of multi-level features Features represents the clue query vector; represents a learnable query; represents the clue key vector; represents the clue value vector; 、 and represents the linear projection weight; represents the cue attention map; express The transpose of represents the dimension of the attention head; represents the softmax activation function; Step 4.3 Calculate the cosine similarity score between the cued attention query and the codebook : ； in, represents cosine similarity; Represents an element in the codebook; A codebook representing a set of feature vectors storing category-related information; Step 4.4 Based on cosine similarity score Perform weighted aggregation on the codebook to generate clue vector : ； in, Indicates the calculation of the number of elements in a collection; Step 4.5: Clue vector Insert Text Feature , obtain enhanced text features : ； in, Represents a splicing operation; Step 4.6 Enhance text features and image depth features Using cross attention mechanism to obtain text-visual fusion features : ； in, Indicates cross-attention; Step 4.7 Fusion of text-visual features Input decoder to generate teacher model prediction results Prediction results with student model .

2. The semi-supervised disguised target detection method based on adaptive data selection and text fusion as claimed in claim 1 is characterized in that The step 1 includes the steps of constructing a disguised target detection dataset RefTextCOD containing image-level directional text: Step 1.1.1 Perform directional text annotation on four mainstream camouflaged target detection datasets: CHAMELEON, CAMO, COD10K, and NC4K. Obtain the precise foreground region of the camouflaged target in the input image based on the dataset ground truth map. Step 1.1.2: Construct corresponding graphic-cue word pairs according to the preset cue word rules; Step 1.1.3 Input the original image, the disguised target foreground area, and the prompt word into the Qwen-VL-Chat model and the GPT4-Vision model respectively to obtain multiple sets of candidate prompt words for each set of images; Step 1.1.4: Manually refine the candidate prompt words to obtain accurate directional text. Step 1.1.5: Organize the results into the format of image-text pairs for the complete dataset.

3. The semi-supervised disguised target detection method based on adaptive data selection and text fusion as claimed in claim 2 is characterized in that The preset prompt word rules are: By giving the original image and using the corresponding segmentation mask to extract the camouflaged target as an auxiliary positioning image, the model is guided by prompt words to gradually identify the foreground of the camouflaged target and its corresponding background physical characteristics, and guide the VLM to aggregate this information to generate the final annotation.

4. The semi-supervised disguised target detection method based on adaptive data selection and text fusion as claimed in claim 1 is characterized in that In step 1, the adaptive data selection module is used to incrementally expand the labeled data set, specifically including: Step 1.2.1 For the first trained model, the adaptive data selection module extracts features in the color, texture, and frequency domains, and uses K-Means clustering to classify the data. The samples close to the cluster center are selected as the initial labeled data, and then proceed to step 2. For the retrained model, the structural similarity index and mean absolute error are calculated for all unlabeled images. Teacher model prediction results after inputting the model Prediction results with student model Data scoring between : ； in, Represents the index of the unlabeled image; and denote the structural similarity index and mean absolute error, respectively; Refers to no mark, superscript Refers to image, subscript Refers to teacher, subscript Refers to student; Step 1.2.2: Standardize the data scores of all unlabeled images, scale the data scores to the numerical range [0, 1] using maximum and minimum normalization, and sort the data according to the normalization results; Step 1.2.3: Select data symmetrically starting from the middle of the sorted results, manually label the selected data, and then add it to the labeled dataset.

5. The semi-supervised disguised target detection method based on adaptive data selection and text fusion according to claim 1, characterized in that: In step 2, the original pixel values ​​of the manually annotated image to be detected are scaled to the interval [0, 1] and normalized using the preset channel mean and variance; the image is then scaled to 640×640, and the semantic segmentation mask corresponding to the image is also scaled to 640×640; a trainable color enhancer is used with a trainable geometric transformation enhancer Perform data augmentation on the scaled image.

6. The semi-supervised disguised target detection method based on adaptive data selection and text fusion according to claim 1, characterized in that: In step 3, the teacher model and student model of the teacher-student framework are both composed of a visual feature extraction backbone network, a text fusion module, and a decoder, and the student model parameters are subjected to exponential moving average to generate the teacher model parameters; the visual feature extraction backbone network uses the Swin pre-trained model to extract image depth features; The decoder uses the BiRefBlock decoder; the visual feature extraction backbone network, text fusion module and decoder all participate in model training.

7. The semi-supervised disguised target detection method based on adaptive data selection and text fusion as claimed in claim 1 is characterized in that In step 5, the total loss of the network The calculation formula is: ； ； ； ； ； ； ； in, represents the supervision loss, which is used to supervise the prediction results of the student model by the real image label; represents a hyperparameter used to balance supervised loss and unsupervised loss; Represents the unsupervised loss, which is used to supervise the prediction results of the teacher model on the prediction results of the student model; represents the supervised segmentation loss, represents the unsupervised segmentation loss, both of which are composed of binary cross entropy loss, intersection-over-union loss, and structural similarity loss; represents the adaptive enhancer loss; represents the binary cross entropy loss; 、 They represent the prediction results of the teacher model and the student model after the labeled image is input into the model; represents the true annotation of the image; represents the loss of the text fusion module; represents the cue attention map; Indicates that the bilinear interpolation algorithm is used to adjust the resolution; represents the mean square error loss; represents the CLIP pre-trained text encoder; The category word that represents the image.

Images