Bursaphelenchus xylophilus disease identification method and system in low-light environment

Abstract

The invention provides a pine wood nematode disease identification method and system in a low-light environment, and belongs to the field of forestry pest and disease monitoring. The method comprises the following steps: acquiring a pine forest aerial image to be detected, and preprocessing the image; processing the preprocessed image by adopting an image enhancement method based on CLIP to obtain an enhanced image; the enhanced image is input into a trained pine wood nematode disease detection model, and feature map extraction is carried out on the enhanced image through an OfficientViT module; performing up-sampling on the feature map by adopting content awareness feature recombination to obtain an enhanced feature map; and inputting the enhanced feature map into the RT-DETR network, and outputting a detection result of the pine wood nematode disease. The method effectively solves the problems of overexposure, underexposure or noise amplification easily occurring under the condition of weak light in a traditional method, improves the detection precision of the model in a low-light environment, and achieves the purpose of meeting the real-time processing requirement of mobile terminal equipment while maintaining the feature extraction capability.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of forestry pest monitoring, and particularly relates to a pine wood nematode disease recognition method and system in a low-light environment. BACKGROUND

[0002] The statements in this section merely provide background information related to the present application and do not necessarily constitute prior art.

[0003] Pine wood nematode disease is a devastating forest disease caused by pine wood nematode, known as "pine cancer". The disease spreads quickly and is difficult to control, posing a serious threat to pine forest resources. Traditional manual patrol methods are inefficient and difficult to implement large-scale monitoring, while image recognition technology based on unmanned aerial vehicle aerial photography provides a new solution for pine wood nematode disease monitoring. However, during the processing of images collected by unmanned aerial vehicles, the light conditions of images taken in different environments are inconsistent, for example, affected by dawn and dusk, tree canopy shading, etc., resulting in poor image quality and blurred features, making images with low light more likely to be misdetected. The commonly used enhancement methods for low-light images (such as RUAS, SCI, Zero-DCE, etc.) either distort the image due to excessive enhancement or affect feature extraction due to noise retention, and cannot effectively improve the quality of pine wood nematode disease detection-related images. The enhancement results are difficult to be effectively utilized by detection models, and the improvement of detection performance is limited. At the same time, existing deep learning models often have large parameter size and high computational complexity, requiring harsh hardware resources, making it difficult to deploy on mobile devices or unmanned aerial vehicle terminals with limited computing power, limiting their practical application in large-scale forest monitoring in the wild. SUMMARY

[0004] To overcome the shortcomings of the prior art, the present application provides a pine wood nematode disease recognition method and system in a low-light environment, which introduces a CLIP-based image enhancement method (CLIP-LIT) to effectively improve the detail representation and brightness distribution of low-light images, reduce overexposed or underexposed areas, and make the features of diseased trees clearer, solving the problem of high false detection rate and missed detection under low light. In order to balance the model accuracy and real-time performance, the RT-DETR (real-time detection Transformer) is used as the core detection network, and the EfficientViT and CARAFE (content-aware feature reorganization) modules are integrated to build a pine wood nematode disease detection model PineSight. This model significantly improves the detection speed and accuracy of pine wood nematode disease while reducing the computational load.

[0005] To achieve the above purpose, one or more embodiments of the present application provide the following technical solutions: The first aspect of the present application provides a pine wood nematode disease recognition method in a low-light environment. A pine wood nematode disease recognition method in a low-light environment, comprising: Obtaining a pine forest aerial image to be detected, and preprocessing the image; An image enhancement method based on CLIP is used to process the preprocessed image to obtain an enhanced image; The enhanced image is input into a trained pine wood nematode disease detection model to output the detection result of the pine wood nematode disease; wherein the trained pine wood nematode disease detection model extracts feature maps from the enhanced image through an EfficientViT module; the feature maps are upsampled using content-aware feature reorganization to obtain enhanced feature maps; the enhanced feature maps are input into an RT-DETR network to output the detection result of the pine wood nematode disease.

[0006] As a further technical solution, the pre-processing of the image is a sliding window method for segmenting the pine forest aerial image.

[0007] As a further technical solution, the image enhancement method based on CLIP is used to process the preprocessed image to obtain an enhanced image, comprising: Initialize the learnable prompt, train the initial enhancement network, and based on the text-image similarity of the CLIP latent space, classify the back light and normal light images through the binary cross entropy loss; Alternately optimize the prompt and the enhancement network parameters through ranking learning to generate the light enhanced image.

[0008] As a further technical solution, the process of training the initial enhancement network includes determining the final loss of the initial enhancement network according to the initialized learning word, using the CLIP-Enhance loss function and the identity loss function.

[0009] As a further technical solution, the process of extracting feature maps from the enhanced image through the EfficientViT module is: The EfficientViT module adopts a sandwich layout, first performs feature segmentation processing on the input feature enhanced image through a group of feedforward neural network layers, enhances the feature representation of the image and extracts local information; The segmented image features are divided into each attention head through a cascaded group of attention layers, the query and key projection channel dimensions of each head are reduced using a redistribution strategy, and the value projection channel dimension is kept consistent with the input dimension; then the processed segmented image features are merged and projected to obtain global information of the global information feature; The global information feature is input into another group of feedforward neural network layers to integrate and enhance the feature representation.

[0010] As a further technical solution, the feature map is up-sampled by using content-aware feature reorganization to obtain an enhanced feature map, including: The content-aware feature reorganization includes a kernel prediction module and a content-aware reorganization module; the kernel prediction module dynamically generates a reorganization kernel according to the content of the feature map to capture the relationship between the features; The content-aware reorganization module is used to perform weighted summation on the features in the local area to up-sample the low-resolution feature map to a high resolution to obtain the enhanced feature map.

[0011] As a further technical solution, the enhanced feature map is input into an RT-DETR network to output the detection result of the pine wilt disease, including: The encoder receives the features output by the backbone network, and the multi-scale features are converted into an image feature sequence through an attention-based internal scale feature interaction module and a cross-scale feature fusion module; The IoU-aware query selects a fixed number of features from the encoder features as the initial target query of the decoder; The decoder iteratively optimizes the initial target query to confirm the position of the diseased tree and generate the final confidence score and bounding box.

[0012] The second aspect of the present application provides a pine wilt disease recognition system in a low-light environment.

[0013] A pine wilt disease recognition system in a low-light environment, including: An image acquisition module configured to acquire a pine forest aerial image to be detected and pre-process the image; An image enhancement module configured to process the pre-processed image by using a CLIP-based image enhancement method to obtain an enhanced image; A detection result output module configured to input the enhanced image into a trained pine wilt disease detection model to output the detection result of the pine wilt disease; wherein the trained pine wilt disease detection model extracts a feature map from the enhanced image through an EfficientViT module; the feature map is up-sampled by using content-aware feature reorganization to obtain an enhanced feature map; and the enhanced feature map is input into an RT-DETR network to output the detection result of the pine wilt disease.

[0014] The third aspect of the present application provides a computer-readable storage medium having a program stored thereon, the program being executed by a processor to implement the steps of the pine wilt disease recognition method in a low-light environment according to the first aspect of the present application.

[0015] The fourth aspect of the present application provides an electronic device, comprising a memory, a processor and a program stored in the memory and executable on the processor, wherein the processor implements the steps of the pine wilt disease recognition method in a low-light environment according to the first aspect of the present application when executing the program.

[0016] The above one or more technical solutions have the following beneficial effects: (1) The image enhancement module (CLIP-LIT) based on the CLIP framework of the present application effectively solves the overexposure, underexposure or noise amplification problems that may occur in weak light conditions by using unsupervised learning and a two-stage optimization strategy. This module can adaptively adjust the light distribution, significantly improve the visibility of disease features in the image, improve the detection accuracy of the model in a low-light environment, and avoid the dependence on manually labeled data.

[0017] (2) The improved RT-DETR architecture of the present application combines the cascaded group attention (CGA) of EfficientViT and the sandwich layout to significantly reduce computational redundancy. The parameter redistribution strategy optimizes the channel width of the key module, compresses the model parameter amount while maintaining the feature extraction capability, reduces the computational amount, and thus meets the real-time processing requirements of mobile terminal devices. Through the CARAFE content-aware upsampling module, an adaptive recombination kernel is dynamically generated to effectively preserve the detailed information of the disease target. Combined with the EfficientViT module, it can accurately capture the representation of pine wilt disease at different scales and significantly improve the recall rate of small target detection.

[0018] The advantages of the additional aspects of the present application will be partially given in the following description, partially become obvious from the following description, or be understood through the practice of the present application. BRIEF DESCRIPTION OF DRAWINGS

[0019] The accompanying drawings, which form a part of the present application, are used to provide further understanding of the present application, and the illustrative embodiments of the present application and their description are used to explain the present application, and do not constitute improper limitations on the present application.

[0020] Figure 1 The method flowchart of the first embodiment.

[0021] Figure 2 The image processing schematic diagram of CLIP-LIT in the first embodiment.

[0022] Figure 3 The structure schematic diagram of the pine wilt disease detection model PineSight constructed in the first embodiment.

[0023] Figure 4 The schematic diagram of the EfficientViT module in the first embodiment.

[0024] Figure 5 A schematic diagram of the CARAFE module in the first embodiment.

[0025] Figure 6 A system structure diagram of the second embodiment. DETAILED DESCRIPTION

[0026] It should be noted that the following detailed description is exemplary in nature and is intended to provide further description of the application. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs.

[0027] It should be noted that the terms used herein are only intended to describe specific embodiments and are not intended to limit exemplary embodiments according to the present application.

[0028] The embodiments in the present application and the features in the embodiments can be combined with each other without conflict.

[0029] Embodiment I The present embodiment discloses a pine wood nematode disease recognition method in a low-light environment; As shown in the figure, a pine wood nematode disease recognition method in a low-light environment includes: Figure 1 Step S1, acquiring a pine forest aerial image to be detected, and pre-processing the image; Step S2, processing the pre-processed image using a CLIP-based image enhancement method to obtain an enhanced image; Step S3, inputting the enhanced image into a trained pine wood nematode disease detection model to output a pine wood nematode disease detection result; wherein the trained pine wood nematode disease detection model extracts feature maps from the enhanced image through an EfficientViT module; the feature maps are up-sampled using content-aware feature reorganization to obtain enhanced feature maps; the enhanced feature maps are input into an RT-DETR network to output a pine wood nematode disease detection result. Specifically, the above steps further include the following contents:

[0030] In step S1, a drone is used to acquire a pine forest aerial image of the detection area, and a sliding window method is used to segment and process the collected image, dividing the image into image blocks with a size of 1024x1024 pixels.

[0031] ​In step S2, by drawing on the idea of ​​Comparative Image Pre-training (CLIP), a CLIP-based image enhancement method (CLIP-LIT) is used to process the preprocessed image. CLIP-LIT can effectively identify the features of backlit images without any paired training data, distinguish heterogeneous regions at different brightness levels, reduce overexposed or underexposed areas, and make the features of diseased trees clearer, thus solving the problem of high false negative and false positive rates under low light conditions. Specifically, as shown... Figure 2 As shown, CLIP-LIT is used to improve the detection performance of underlit samples in the dataset. In this embodiment, the CLIP-LIT framework mainly consists of two stages. The first stage learns a set of initial cue word pairs to roughly distinguish between backlit and normally lit photos and trains a preliminary enhancement network. The second stage introduces ranking learning to obtain normally lit images, enhancement results, and backlit images, and iteratively optimizes the learned cue words and enhancement network accordingly.

[0032] Furthermore, in the first stage, CLIP-LIT initializes learnable cue words to roughly describe backlit and well-lit images and trains the initial augmentation network. Given a backlit image and a well-lit image, the model randomly initializes a positive cue word and a negative cue word. Image encoder and text encoder The latent codes for the image and the cue word pairs are obtained separately. Based on the text-image similarity in the CLIP latent space, the model uses a binary cross-entropy loss function to classify backlit and well-lit images, thereby learning the initial cue word pairs:

[0033]

[0034] in, The binary cross-entropy loss function; The label for the current image; A function for measuring the similarity between images and text; As a prompt word, , n Indicates positive. p Indicates negative; 0 indicates a negative sample 1 indicates a positive sample .

[0035] Based on the initial prompts, the augmentation network is further trained using the CLIP-Enhance loss function and the identity loss function. The CLIP-Enhance loss function measures the similarity between the augmented result and the prompts. The identity loss function encourages the augmented result to be similar to the backlit image in content and structure. The final loss for training the augmentation network is a combination of these two losses.

[0036]

[0037]

[0038] in, The CLIP-Enhance loss function; For the enhanced image; Negative indicator words; For identity loss function; Indicates the image encoder's first... Layer weights It is a coefficient used to balance the weights of different loss terms.

[0039] Augmented network for estimating illumination maps and through Obtain initial enhancement results.

[0040] In the second stage, the cue words and enhancement network parameters are alternately optimized through sequential learning to improve the accuracy of the learned cues in distinguishing backlit images, enhanced images, and well-lit images, as well as perceiving heterogeneous regions with different brightness levels. Based on the results of the current enhancement network enhancement, the negative similarity score between the cue pair and the image is used. and edge sorting loss Update the prompt. The formula is as follows:

[0041]

[0042] in, This indicates the score difference between the bright or enhanced result and the backlit image. (Settings) The value is set to 0.9 to increase the distance between the backlit image and the illuminated image. This indicates the score difference between the enhanced result and the brightened image. (Settings) Set it to 0.2 to ensure the enhancement result is close to the brightened image.

[0043] To ensure performance improvements in each iteration, previous enhancement results obtained from the previous round of network enhancement are retained. This was then used for ranking. Two sets of enhanced results. and The constraint is incorporated, so that the newly learned prompt pays more attention to the light and color distribution of the image. The loss function becomes:

[0044] where, represents the gap between the newly enhanced result and the previous enhanced result. By setting as the standard, the similarity of the two groups of images is maintained. By adjusting the enhancement network, the process is the same as the first stage. The CLIP-Enhance loss is calculated using the refined prompt, that is, , and the enhanced training data is generated from the updated network.

[0045] In summary, in the second stage, in each iteration, the enhanced result (the initial enhanced result obtained in the first stage at the beginning), the previous enhanced result , the backlit image and the normally lit image are used to update the prompt and . By edge ranking loss and , it is ensured that the enhanced result is closer to the representation of the normally lit image , away from the representation of the backlit image , and maintains the similarity with . Then, using the refined prompt, the enhanced image , ) and the backlit image are used to train the enhancement network again through the CLIP-Enhance loss function and the identity loss function of the first stage, to generate a new enhanced image .

[0046] In step S3, the trained pine wilt disease detection model PineSight is used to extract feature maps from the enhanced image through the EfficientViT module; wherein the structure of the trained pine wilt disease detection model PineSight is as shown in Figure 3 . The content-aware feature reorganization is used to upsample the feature maps to obtain enhanced feature maps; the enhanced feature maps are input into the RT-DETR network to output the detection result of the pine wilt disease.

[0047] Specifically, in step S31, the EfficientViT module is composed of a sandwich layout, a cascade group attention (CGA), and a parameter redistribution strategy, as shown in Figure 4The EfficientViT module first processes the input features through a set of feed-forward network layers (FFN) to enhance feature representation and extract local information. Then the segmented features are divided into each head through a set of attention layers, and the processed segmented features are merged and projected to obtain the global information of the features. Finally, another set of FFN layers is used to process the features after the attention mechanism to further integrate and enhance the feature representation. The sandwich layout reduces the number of memory-constrained self-attention layers and increases the memory-efficient FFN layers, thereby speeding up the inference. CGA divides different features into attention heads, reducing the computational redundancy in multi-head attention, increasing the diversity of attention maps, and enabling the improvement of network depth without adding extra parameters. The parameter redistribution strategy improves parameter efficiency by adjusting the channel width of key and secondary modules, prevents the loss of feature information in key modules, and speeds up inference by eliminating redundant parameters in secondary modules. EfficientViT significantly improves the computational efficiency of the model while maintaining high performance, achieving model compression and efficiency optimization, easy deployment, and meeting the needs of real-time and rapid detection of pine wilt disease.

[0048] Specifically, a single self-attention layer is sandwiched between multiple FFN layers to form a sandwich layout. In addition, an additional token interaction layer is applied before each FFN layer, and the interaction layer uses deep convolution DWConv to introduce an inductive bias of local information. is the th block, i.e., the complete input features of the image after the S2 step. Before and after the single self-attention layer, the module applies N FFNs to convert it to . This layout reduces the memory consumption caused by self-attention layers in the model and enhances the communication between feature channels by applying more FFN layers, thereby improving the model capacity. The calculation formula is as follows.

[0049]

[0050] Further, the cascaded group attention (CGA) is used to provide feature segmentation for each head, thereby explicitly decomposing the attention calculation between heads to reduce the computational burden. In addition, it also encourages each layer to learn a feature projection containing rich information. CGA calculates the attention map for each head in a cascaded manner, further enhancing the model capacity, and the specific structure is shown in the figure. The formula for calculating attention is as follows.

[0051]

[0052]

[0053]

[0054] where, is the input feature is the first th partition of the input feature, where , is the total number of heads. The projection layer maps the partitions of the input feature to different subspaces. The linear layer reprojects the concatenated output features back to the dimension consistent with the input. is the sum of the input feature partitions and the th head output, which replaces as the new input feature to the th head.

[0055] The cascading method accumulates the output of each head to the next head, gradually improving the feature representation. In addition, after Q projection, another token interaction layer is used to assist the self-attention mechanism, together capturing local and global relationships, further enhancing the feature representation.

[0056] To improve parameter efficiency, EfficientViT proposes a redistribution strategy based on Taylor importance analysis. This strategy expands the channel width of key modules while reducing less important modules, redistributing parameters within the network to reduce computational redundancy. The Q and K projections of each head in all stages are set to a smaller channel dimension. For V projection, it is set to the same dimension as the input embedding. The expansion of the feed-forward network (FFN) is reduced from 4 times to 2 times. Important modules are allocated more channels to focus on feature information, while redundant parameters in less important modules are removed, thereby improving model efficiency.

[0057] In step S32, the feature map is upsampled by content-aware feature reorganization to obtain an enhanced feature map. Specifically, content-aware feature reorganization (CARAFE) is a general, lightweight and effective up-sampling operation. CARAFE aggregates context information in a larger receptive field and generates an adaptive kernel in real time, while only requiring minimal computational overhead. Therefore, by introducing a CARAFE up-sampling module to replace the traditional nearest neighbor interpolation up-sampling method, the feature information can be better captured. CARAFE includes a kernel prediction module and a content-aware reorganization module.

[0058] The kernel prediction module is responsible for dynamically generating a reorganization kernel according to the content of the feature map, which can better capture the relationship between features and achieve lightweight and efficient computation. This module consists of a channel compressor, a content encoder, and a kernel normalizer. Given a feature map of size , and an integer up-sampling ratio , there is an original position and a sub-region centered at position . That is, the neighborhood of , i.e. pixels at the specific position of the original feature map. The channel compressor uses a convolutional layer with kernel size to reduce the input feature channels from to . The content encoder uses a convolutional layer with kernel size to generate a reassembly kernel, where , is the size of the reassembly kernel. Each target position on requires a reassembly kernel with size . Therefore, the final output size of this module is . The kernel normalizer uses the softmax function to normalize the reassembly kernel , ensuring that the sum of kernel values is 1.

[0059] In the output , there is a target coordinate , where . Finally, the kernel prediction module predicts a position kernel for each position , as shown in the formula.

[0060]

[0061] The content-aware reassembly module performs weighted summation of features within a local region, up-samples the low-resolution feature map to high resolution while preserving rich semantic information, ultimately improving the accuracy of image processing tasks. For each reassembly kernel , the module will perform feature reassembly through the function . For the target position and its corresponding square region , the formula of this function is as follows. In each pixel pair contributes differently to the up-sampled pixel . The semantic of the reassembled feature map can be stronger than the original feature map, as the information of related points within the local region can get more attention.

[0062]

[0063] where is the up-sampled pixel of the newly generated feature map.

[0064] Step S33, input the enhanced feature map into the RT-DETR network, and output the detection result of the pine wilt disease.

[0065] RT-DETR (Real-Time Detection Transformer) is an end-to-end detector that introduces the Transformer architecture into the real-time object detection task. This model does not require non-maximum suppression (NMS) and significantly improves the detection speed while maintaining high accuracy. RT-DETR is composed of a backbone network, a hybrid encoder, and a Transformer decoder equipped with an auxiliary prediction head. The encoder receives the features output by the backbone network and converts multi-scale features into image feature sequences through the attention-based internal scale feature interaction module (AIFI) and the cross-scale feature fusion module (CCFM). Then, the IoU-aware query selection selects a fixed number of features from the encoder features as the initial target queries for the decoder. The decoder with an auxiliary prediction head iteratively optimizes these queries to generate classes and bounding boxes. Again, after IoU-aware query selection, a fixed number of features are sampled from the encoder output as initial target queries, and finally, the decoder iteratively refines these queries to predict confidence scores and bounding boxes. By generating bounding boxes, the location of diseased trees and the area occupied by the diseased trees can be obtained.

[0066] To verify the effectiveness of the CLIP-LIT image enhancement method in the present application, three weak light image enhancement technologies, RUAS, SCI, and Zero-DCE, were introduced for comparison. The experimental results are shown in Table 1. The CLIP-LIT method used in the present application showed significant advantages in core indicators such as precision and F1 score, with an mAP@0.5 value of 83.11%, surpassing other comparison methods.

[0067] Table 1. Comparison of detection results of the present application and other weak light image enhancement technologies

[0068] To more intuitively demonstrate the overall detection performance of the present application, the detection results of the present application are compared with the experimental results of conventional object detection algorithms such as Faster-RCNN, Retinanet, YOLOv5, YOLOX and DETR. The specific data are shown in Table 2. As can be seen from the table, PineSight is significantly better than other models in the mAP@0.5 and mAP@0.5:0.95 indicators. This shows that the model can maintain good performance under different environmental conditions, showing strong robustness and adaptability. In terms of accuracy, the precision and recall rate of the present application in detecting pine nematode disease can reach 91.08% and 83.38% respectively. Therefore, the PineSight model constructed by the present application can not only accurately identify diseased trees, but also effectively distinguish healthy areas from diseased areas, while ensuring the recall rate of the model.

[0069] Table 2. Detection results of different deep learning algorithms

[0070] Further, in order to further analyze the interaction between the introduced modules, an ablation experiment is also performed to specifically evaluate the influence of each algorithm on the model performance, and the experimental results are shown in Table 3. It can be seen that the CLIP-LIT module optimizes the quality of the weak light image; the EfficientViT module improves the calculation efficiency and reduces redundant operations; and the CARAFE module further enhances the detection capability through feature reorganization technology. The model achieves high levels in precision (91.08%) and recall (83.38%), and the F1 score (87.06%) further proves that the balance between precision and recall is excellent. This shows that PineSight can achieve good results in the detection of datasets under different light conditions.

[0071] Table 3. Detection results obtained by ablation experiment

[0072] Example Two The present embodiment discloses a pine nematode disease recognition system in a low-light environment; As Figure 6 shown, a pine nematode disease recognition system in a low-light environment includes: An image acquisition module configured to acquire a pine forest aerial image to be detected and pre-process the image; An image enhancement module configured to process the pre-processed image using a CLIP-based image enhancement method to obtain an enhanced image; The detection result output module is configured to input the enhanced image into the trained pine wood nematode disease detection model to output a detection result of the pine wood nematode disease; wherein the trained pine wood nematode disease detection model extracts a feature map from the enhanced image through the EfficientViT module; performs content-aware feature reorganization on the feature map to obtain an enhanced feature map; and inputs the enhanced feature map into the RT-DETR network to output the detection result of the pine wood nematode disease.

[0073] Embodiment three The purpose of this embodiment is to provide a computer-readable storage medium.

[0074] A computer-readable storage medium having stored thereon a computer program, which, when executed by a processor, implements the steps of the pine wood nematode disease recognition method in a low-light environment according to embodiment 1.

[0075] Embodiment four The purpose of this embodiment is to provide an electronic device.

[0076] An electronic device comprising a memory, a processor, and a program stored on the memory and executable on the processor, wherein the processor implements the steps of the pine wood nematode disease recognition method in a low-light environment according to embodiment 1 when executing the program.

[0077] The steps and methods involved in the above embodiments two, three, and four correspond to embodiment one, and the specific implementation can be referred to the relevant description part of embodiment one. The term "computer-readable storage medium" should be understood to include a single medium or multiple media of one or more instruction sets; it should also be understood to include any medium that can store, encode, or carry instruction sets for execution by a processor and cause the processor to perform any method in the present application.

[0078] Those skilled in the art should understand that the above-mentioned modules or steps of the present application can be realized by a general computer device, alternatively, they can be realized by program codes executable by a computing device, so that they can be stored in a storage device for execution by a computing device, or they can be made into individual integrated circuit modules, or a plurality of modules or steps among them can be made into a single integrated circuit module. The present application is not limited to any specific combination of hardware and software.

[0079] The above describes the specific embodiments of the present application in conjunction with the accompanying drawings, but is not a limitation on the scope of protection of the present application. Those skilled in the art should understand that various modifications or variations made by those skilled in the art on the basis of the technical solutions of the present application without creative labor are still within the scope of protection of the present application.

Claims

1. A method for identifying pine wilt disease in low-light environments, characterized in that, include: Acquire aerial images of the pine forest to be detected and preprocess the images; The preprocessed image is processed using a CLIP-based image enhancement method to obtain the enhanced image; The enhanced image is input into a trained pine wilt disease detection model, which outputs the detection result of pine wilt disease. The trained pine wilt disease detection model extracts feature maps from the enhanced image using the EfficientViT module. Content-aware feature reconstruction is used to upsample the feature maps to obtain enhanced feature maps. The enhanced feature maps are then input into the RT-DETR network to output the detection result of pine wilt disease.

2. The method for identifying pine wilt disease in a low-light environment as described in claim 1, characterized in that, The image preprocessing process involves segmenting the aerial image of the pine forest using a sliding window method.

3. The method for identifying pine wilt disease in a low-light environment as described in claim 1, characterized in that, The preprocessed image is processed using a CLIP-based image enhancement method to obtain the enhanced image, including: Initialize learnable cue words, train an initial augmentation network, and classify backlit and normal-lit images using text-image similarity based on CLIP latent space and binary cross-entropy loss. By sequentially learning and alternately optimizing prompts and enhancing network parameters, an image with enhanced illumination is generated.

4. The method for identifying pine wilt disease in a low-light environment as described in claim 3, characterized in that, The process of training the initial augmentation network includes determining the final loss of the initial augmentation network based on the initial learned words using the CLIP-Enhance loss function and the identity loss function.

5. The method for identifying pine wilt disease in a low-light environment as described in claim 1, characterized in that, The process of extracting feature maps from the enhanced image using the EfficientViT module is as follows: The EfficientViT module adopts a sandwich layout. First, it performs feature segmentation processing on the input feature-enhanced image through a set of feedforward neural network layers to enhance the feature representation of the image and extract local information. The segmented image features are distributed to each attention head through a cascaded attention layer. A redistribution strategy is used to reduce the query and key projection channel dimensions of each head while keeping the value projection channel dimensions consistent with the input dimensions. The processed segmented image features are then merged and projected to obtain the global information of the global information features. The global information features are input into another set of feedforward neural network layers to integrate and enhance the feature representation.

6. The method for identifying pine wilt disease in a low-light environment as described in claim 1, characterized in that, Content-aware feature reconstruction is used to upsample the feature map, resulting in an enhanced feature map, including: The content-aware feature reconstruction includes a kernel prediction module and a content-aware reconstruction module; the kernel prediction module dynamically generates a reconstruction kernel based on the content of the feature map and captures the relationships between features. The content-aware reconstruction module is used to perform weighted summation of features in a local area, upsampling the low-resolution feature map to a high resolution to obtain an enhanced feature map.

7. The method for identifying pine wilt disease in a low-light environment as described in claim 1, characterized in that, The enhanced feature map is input into the RT-DETR network, which outputs the detection results of pine wilt disease, including: The encoder receives the features output by the backbone network and transforms the multi-scale features into an image feature sequence through an attention-based internal scale feature interaction module and a cross-scale feature fusion module. The IoU-aware query selects a fixed number of features from the encoder features as the initial target query for the decoder. The decoder is used to iteratively optimize the initial target query, confirm the location of diseased trees, and generate the final confidence score and bounding box.

8. A system for identifying pine wilt disease in low-light environments, characterized in that: include: The image acquisition module is configured to acquire aerial images of the pine forest to be detected and to preprocess the images. The image enhancement module is configured to process the preprocessed image using a CLIP-based image enhancement method to obtain the enhanced image. The detection result output module is configured to: input the enhanced image into the trained pine wilt disease detection model and output the detection result of pine wilt disease; wherein the trained pine wilt disease detection model extracts feature maps from the enhanced image through the EfficientViT module; upsamples the feature maps using content-aware feature reconstruction to obtain enhanced feature maps; and inputs the enhanced feature maps into the RT-DETR network to output the detection result of pine wilt disease.

9. A computer-readable storage medium having a program stored thereon, characterized in that, When executed by a processor, the program implements the steps in the method for identifying pine wilt disease in a low-light environment as described in any one of claims 1-7.

10. An electronic device comprising a memory, a processor, and a program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps in the method for identifying pine wilt disease in a low-light environment as described in any one of claims 1-7.

Images