Fan blade fault image detection method in weak light environment

Abstract

The invention discloses a fan blade fault image detection method in a weak light environment, and the method comprises the steps: carrying out the enhancement processing of a fan blade image collected under a weak light condition through a Bi-Enhancer model, and the Bi-Enhancer comprises a reflection convolution layer, a lightweight feature aggregation module, a high-level feature enhancement module, a double-layer feature fusion module, and an image iteration optimization module; and inputting the enhanced weak light image into a target detection model for processing to obtain a fault detection result. The Bi-Enhancer provided by the invention has an efficient feature fusion and enhancement mechanism, and can realize better performance and higher processing speed in an advanced vision task; the Bi-Enhancer not only is excellent in performance in a weak light environment, but also has a good generalization ability; therefore, the method can better deal with a fan blade fault detection task in a weak light environment, and has higher detection precision and detection efficiency compared with an existing method.

Description

Technical Field

[0001] The present invention relates to the fields of machine vision, deep learning, image feature aggregation and enhancement, wind farm operation and maintenance technology, etc., and particularly relates to a method for detecting fault images of wind turbine blades. Background Art

[0002] At present, the blade inspection in the wind power industry mainly adopts the method of "manual + gondola", but this method has problems such as low efficiency, the need for shutdown inspection, being greatly interfered by subjective factors, and being prone to safety accidents. Therefore, in recent years, researchers at home and abroad have gradually developed a variety of online detection technologies, such as fiber Bragg grating, infrared imaging, acoustic emission, vibration signal and non-contact acoustics, etc. With the rapid development of computer vision and unmanned aerial vehicle (UAV) inspection technology, using UAVs for visual inspection provides a new means for blade damage detection. This method can not only avoid safety accidents that may occur during manual inspection, but also reduce the influence of human subjective factors during the inspection process. At the same time, compared with technologies such as acoustic emission, vibration and non-contact acoustics, visual inspection can accurately identify the type, location and degree of defects on the blade surface, and is particularly more sensitive in the detection of early damage (such as cracks, lightning strikes, etc.). However, due to the change of light conditions in the natural environment, when using computer vision to detect blade surface defects, it is susceptible to the influence of time and weather. In images taken under low-light and dark-light conditions, the contrast of wind turbine blade images is significantly reduced, the defect features are blurred, which easily leads to misjudgment or missed detection.

[0003] The reason for the blurred defect features on the blade surface is that the wind turbine unit is long-term exposed to the natural environment, and the acquisition of blade images is directly affected by light and meteorological conditions. Images collected in cloudy days, backlight and dusk usually have problems such as low visibility, poor contrast and large measurement noise, resulting in difficulty in identifying the features of the defect area. At present, although the camera comes with a light enhancement function and performs well in brightness improvement, color balance and visual effect, it often causes problems such as loss of texture information and smoothing of edge details.

[0004] At present, a large number of relevant literatures at home and abroad have studied the target detection model in low-light environments and proposed effective methods for image enhancement target detection in low-light and dark-light environments.

[0005] 1. In the article titled "Multi-modal Fusion Target Detection Method under Weak Light Conditions of Unmanned Aerial Vehicles", the authors proposed a multi-scale differential attention fusion detection method that couples illumination conditions and contrast. First, a multi-scale differential attention fusion detection network guided by information perception was designed. By calculating the illumination information of the image and the local contrast of the target through the information perception module, it guided the multi-scale differential attention module to perform deep cross-fusion on the intra-modal and inter-modal features of visible light and infrared images, so as to improve the detection and recognition accuracy of unmanned aerial vehicles for ground targets under weak light conditions. Although this method can significantly improve the detection performance of unmanned aerial vehicles for targets under weak light conditions and has good real-time performance, it is difficult to construct and open-source the target detection dataset of unmanned aerial vehicles under weak light and dark light conditions, and the process is cumbersome and complex.

[0006] 2. Liu Susu et al. used the average brightness index as a threshold in the patent application with the application number CN202311542442.4 and the title "A Multi-classification Detection Method for Saw Chain Defects Based on Adaptive Enhancement of Weak Light Scene Images" to determine whether the collected image needs to be enhanced under weak light. If it is lower than the threshold, the collected image is input into the RRDNet network for adaptive enhancement. If it is greater than or equal to the threshold, the collected image uses the saw chain part segmentation algorithm to complete part segmentation, segment the graph and form a defect detection queue, establish a dataset, and based on the improved ResNet34 network model, achieve multi-classification of saw chain defects under weak light scenes. Although this method uses the RRDNet network to perform adaptive enhancement on low-brightness images, effectively improving the contrast and detail performance of images under weak light scenes and enhancing the accuracy of defect detection. However, using the average brightness index as a threshold has certain limitations and cannot flexibly cope with different illumination changes or complex environments, which may lead to misjudgment or poor enhancement effects.

[0007] 3. In the article titled "Efficient network architecture for target detection in challenging low-light environments", the authors Qiang Liu et al. proposed an improved target detection model aiming to solve target detection under dark light conditions. The image enhancement technology designed for dark light environments focuses on the restoration of preprocessed images to achieve real and natural visual quality. The main improvements include introducing a multi-layer fine-grained feature prediction network architecture, realizing hierarchical feature fusion and fine feature extraction, thereby improving the accuracy of the model, reducing the number of parameters, and enhancing the precision of the target area through a dynamic detection output head, ultimately improving the overall detection accuracy. This method relies on specific image preprocessing strategies during the low-light image enhancement process. When facing extremely dark light, weak light, or high-contrast scenes, the enhancement effect is insufficient and it cannot fully adapt to different complex environments.

[0008] Meanwhile, target detection algorithms based on low light are less applied in the detection of fan blades. Moreover, the types of surface defect detection targets of fan blades are diverse, the scale changes greatly, and they are greatly affected by natural light. Therefore, there is still much room for improvement in the existing low light detection algorithms for fan blade target detection. Summary of the Invention

[0009] In view of this, the purpose of the present invention is to provide a method for detecting fault images of fan blades in low light environments, so as to solve the technical problems of improving the detection accuracy and efficiency of fan blade faults in low light environments.

[0010] The method for detecting fault images of fan blades in low light environments according to the present invention includes the following steps:

[0011] I) Enhance the fan blade image collected under low light conditions through the Bi-Enhancer model. The Bi-Enhancer model includes a reflection convolution layer, a lightweight feature aggregation module, a high-level feature enhancement module, a double-layer feature fusion module, and an image iterative optimization module. The enhancement process includes:

[0012] 1) Input the original low light image of the fan blade into the reflection convolution for processing to generate a high-level image;

[0013] 2) Input the original low light image into the lightweight feature aggregation module for processing to generate a low-level feature map; input the high-level image into the lightweight feature aggregation module for processing to generate a high-level feature map;

[0014] 3) Input the high-level feature map into the high-level feature enhancement module for processing to generate a high-level enhanced feature map;

[0015] 4) Input the high-level image, the high-level enhanced feature map, and the low-level feature map into the double-layer feature fusion module for processing to generate a feature fusion enhanced map;

[0016] 5) Input the original low light image of the fan blade, the high-level enhanced feature map, and the feature fusion enhanced map into the image iterative optimization module for processing to generate an enhanced low light image;

[0017] II) Input the enhanced low light image obtained in step 5) into the target detection model for processing to obtain the fan blade fault detection result.

[0018] Furthermore, in step 1), the process of generating a high-level image by the reflection convolution is as follows:

[0019] I u = RefConv(I)

[0020] where: RefConv represents the reflection convolution operation; I is the original low light image of the fan blade, Iu It is the generated high-level image.

[0021] Furthermore, in step 2), the process of the lightweight feature aggregation module for processing the input image includes:

[0022] F1 = SiLU[Conv(I in )]

[0023]

[0024] F3 = SiLU[Conv(F2)]

[0025]

[0026] F5 = SiLU[Conv(F4)]

[0027]

[0028] F out = SiLU[Conv(F6)]

[0029] where: Conv is the conventional convolution operation, SiLU is the activation function, is the concatenation of features along the channel direction, I in is the input image, F out is the output image.

[0030] Furthermore, in step 3), the process of the high-level feature enhancement module for processing the input high-level feature map includes:

[0031] F s = FuseConv(F u )

[0032] F sr = FuseConv(F s )

[0033] F d = DWConv(F sr )

[0034] F u ' = DWConv(F d )

[0035] where, F u is the high-level feature map, F u ' is the high-level enhanced feature map;

[0036] FuseConv(F u ) = SiLU[BatchNorm(C1onv(F u)) + BatchNorm(C2onv(F u ))]

[0037] FuseConv(F s ) = SiLU[BatchNorm(C1onv(F s )) + BatchNorm(C2onv(F s ))]

[0038] DWConv(F sr ) = SiLU[Point_C2onv(Depth_C1onv(F sr ))]

[0039] DWConv(F d ) = SiLU[Point_C2onv(Depth_C1onv(F d ))]

[0040] Among them, F u is the high-level feature map, F u ' is the high-level enhanced feature map, BatchNorm is the regularization operation, C1onv and C2onv are conventional convolution operations, Depth_C1onv is the depth convolution operation, Point_C2onv is the pointwise convolution operation, and SiLU is the activation function.

[0041] Furthermore, in step 4), the process of the double-layer feature fusion module processing the input image includes:

[0042] a) Processing the low-level feature map and the high-level image through the reflection convolution to align the space and channels of the feature map output by the reflection convolution with the high-level enhanced feature map:

[0043] F b ' = RefConv(F b )

[0044] F i = RefConv(I u )

[0045] Among them, RefConv is the reflection convolution operation, F b is the low-level feature map, and I u is the high-level image;

[0046] b) Splitting the feature maps F′ b , F i and F u ' into N sub-feature maps respectively and adding a new dimension T. The process is as follows:

[0047]

[0048] Among them, n represents the decomposition step, n = 1, 2, 3, …, N; c is the number of feature fusion channels;

[0049] c) Fuse the sub-feature maps through N fusion sub-modules, and the fusion process of each fusion sub-module is as follows:

[0050] First, splice the three sub-feature maps and along the direction of the penultimate dimension. H is the height of the original low-light image, W is the width of the original low-light image, and C is the number of channels of the original low-light image, to obtain the feature map Subsequently, multiply the three feature sub-maps with element-wise to obtain three cross-attention weights and The element-wise multiplication operation process is as follows:

[0051]

[0052] Among them, ⊙ is the element-wise multiplication operation. Subsequently, splice and sum all the cross-attention weights along the direction of dimension S, and then perform feature normalization along the last dimension to obtain the double-layer attention weight of a single sub-module The calculation process of the attention weight is as follows:

[0053]

[0054] Among them, is splicing along the direction of dimension S;

[0055] d) Fuse the double-layer attention weight with the three sub-feature maps without adding dimension T F i n , through weighted summation to fuse on the grouped dimension of the sub-module, to obtain a more informative feature map The weighted summation process is as follows:

[0056]

[0057] e) Use bilinear interpolation to upsample F bi by 8 times to obtain the enhanced feature map Meanwhile, in order to effectively reduce the loss of underlying information, perform a skip connection between F′ e and the underlying feature map F b to obtain the feature fusion enhanced map The above process is as follows:

[0058] F′e = Up(F bi ), sf = 8

[0059] F e = F′ e + F b

[0060] Among them, Up is bilinear interpolation upsampling, and sf is the upsampling scale factor.

[0061] Furthermore, in step 5), the processing of the input image by the iterative optimization module includes:

[0062] 1) Concatenate the feature fusion enhanced map and the high-level enhanced feature map in the channel direction to obtain a mixed feature map Subsequently, use bilinear interpolation to upsample F fu by 8 times to obtain a more informative iterative feature map The feature concatenation and upsampling processes are as follows:

[0063]

[0064] F i = Up(F fu ), sf = 8

[0065] 2) Decompose the iterative feature map to obtain iterative factors, including: Expand the iterative feature map F i in the channel direction and decompose it into 2N iterative factors F n H×W×3 for image iterative optimization. Each iterative factor corresponds to a different feature representation. The feature splitting process is as follows:

[0066]

[0067] 3) Iterative optimization: Apply the iterative factors to the input original low-light image I for e times of iterative optimization to generate an enhanced low-light image I e , and the iterative optimization process is as follows:

[0068]

[0069] When n = 1, I0 is the input original low-light image, and I e is the enhanced low-light image output after iterative optimization.

[0070] Furthermore, the Bi-Enhancer model and the object detection model are cascaded, and the output of the Bi-Enhancer model is directly fed into the backbone network of the detector model; the Bi-Enhancer model and the object detection model are trained in an end-to-end manner. During the training process, the Bi-Enhancer model is implicitly optimized through the loss function of the object detection model without setting a separate loss function.

[0071] Advantages of the present invention:

[0072] 1. For the method for detecting faults in fan blades under low-light conditions, the proposed Bi-Enhancer model has a simplified network structure, which has an efficient feature fusion and enhancement mechanism, and can achieve better performance and faster processing speed in advanced vision tasks; the Bi-Enhancer not only performs well in low-light environments but also has good generalization ability; this enables the method of the present invention to better handle the task of detecting faults in fan blades under low-light conditions and has higher detection accuracy and detection efficiency compared with existing methods.

[0073] 2. Currently, low-light image enhancement methods focus too much on optimizing human visual perception and relatively neglect improving the machine readability of images, and rely too much on supervised training of synthetic data pairs, resulting in low generalization ability for real-scene images. To address the above problems, from the perspective of machine readability, the method of the present invention conducts cascaded end-to-end joint training on the Bi-Enhancer model and the object detection model. The training of the Bi-Enhancer model does not consider the loss function. By optimizing the feature expression of the image, the Bi-Enhancer model not only improves the detection accuracy of fan blade targets in low-light and dark environments but also maintains the detection efficiency. Description of the drawings

[0074] Figure 1 It is the overall architecture diagram of the Bi-Enhancer model.

[0075] Figure 2 It is the structural diagram of the lightweight feature aggregation module.

[0076] Figure 3 It is the structural diagram of the high-level feature enhancement module.

[0077] Figure 4 It is the structural diagram of the fusion convolution.

[0078] Figure 5 It is the structural diagram of the depthwise separable convolution.

[0079] Figure 6 It is the structural diagram of the double-layer feature fusion module. Detailed implementation manners

[0080] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0081] The method for detecting the fault image of the fan blade in low-light environment in this embodiment includes the following steps:

[0082] I) Enhance the fan blade image collected under low-light conditions through the Bi-Enhancer model. The Bi-Enhancer model includes a reflection convolution layer, a lightweight feature aggregation module (Lightweight Feature Aggregation Module, LFAM), a high-level feature enhancement module (High-level Feature Enhancement Module, SFEM), a bi-level feature fusion module (Bi-level Feature Fusion Module, BFFM), and an image iterative optimization module (Image Iterative Optimization Module, IIOM). The overall architecture of the Bi-Enhancer model is as Figure 1 shown. The enhancement process includes:

[0083] 1) Input the original low-light image of the fan blade into the reflection convolution for processing to generate a high-level image. The process of generating a high-level image by the reflection convolution is as follows:

[0084] I u = RefConv(I)

[0085] where: RefConv represents the reflection convolution operation, the convolution kernel size k = 9, the convolution kernel stride s = 8, and the reflection padding size p = 4; I is the original low-light image of the fan blade, H is the height of the original low-light image, W is the width of the original low-light image, I u is the generated high-level image,

[0086] 2) Input the original low-light image into the lightweight feature aggregation module for processing to generate a low-level feature map Input the high-level image into the lightweight feature aggregation module for processing to generate a high-level feature map The number of channels C of the low-level feature map and the high-level feature map takes the value of 24. In this step, the process of the lightweight feature aggregation module for processing the input image includes:

[0087] F1 = SiLU[Conv(I)], k = 1, s = 1, p = 0

[0088]

[0089] F3 = SiLU[Conv(F2)], k = 1, s = 1, p = 0

[0090]

[0091] F5 = SiLU[Conv(F4)], k = 1, s = 1, p = 0

[0092]

[0093] F b = SiLU[Conv(F6)], k = 3, s = 1, p = 1

[0094] Where: Conv is the conventional convolution operation, k is the convolution kernel size, s is the convolution kernel stride, p is the padding size; SiLU is the activation function, is the concatenation of features along the channel direction; I in is the input image, I in specifically the original low-light image I and the high-level image I u ; F out is the output image.

[0095] 3) High-level feature map contains rich abstract semantic information, which is crucial for enhancing the model's perception ability, understanding ability, and the performance of high-level vision tasks. However, while enhancing these semantic information, it often leads to the loss of detailed information such as texture and contour. To solve this problem, this embodiment adopts a new high-level feature enhancement module (HFEM), which includes two groups of fused convolutions (FuseConv) for increasing the number of channels of the high-level feature map F u to enhance the expressive ability of features; it also includes two groups of depthwise separable convolutions (DWConv) for reducing the number of channels of the feature map to ensure that the number of channels of the finally output high-level enhanced feature map F u ' remains unchanged. Input the high-level feature map into the high-level feature enhancement module for processing to generate the high-level enhanced feature map; the processing of the high-level feature map input by the high-level feature enhancement module includes:

[0096] F s = FuseConv(F u )

[0097] F sr = FuseConv(F s )

[0098] F d = DWConv(F sr )

[0099] F u' = DWConv(F d )

[0100] where F u is the high-level feature map, and F u ' is the high-level enhanced feature map;

[0101] FuseConv(F u ) = SiLU[BatchNorm(C1onv(F u )) + BatchNorm(C2onv(F u ))]

[0102] FuseConv(F s ) = SiLU[BatchNorm(C1onv(F s )) + BatchNorm(C2onv(F s ))]

[0103] DWConv(F sr ) = SiLU[Point_C2onv(Depth_C1onv(F sr ))]

[0104] DWConv(F d ) = SiLU[Point_C2onv(Depth_C1onv(F d ))]

[0105] where F u is the high-level feature map, and F u ' is the high-level enhanced feature map, BatchNorm is a regularization operation, C1onv and C2onv are conventional convolution operations, the convolution kernel size k of C1onv is 1, the convolution kernel stride s is 1, and the padding size p is 0. The convolution kernel size k of C2onv is 3, the convolution kernel stride s is 1, and the padding size p is 1; Depth_C1onv is a depth convolution operation, the convolution kernel size k of Depth_C1onv is 3, the convolution kernel stride s is 1, and the padding size p is 0; Point_C2onv is a pointwise convolution operation, the convolution kernel size k of Point_C2onv is 1, the convolution kernel stride s is 1, and the padding size p is 0; SiLU is an activation function.

[0106] 4) In the image feature map, feature maps at different levels have their own characteristics. High-level feature maps contain rich image semantic information, low-level feature maps retain a large amount of detailed information, and middle-level feature maps incorporate a certain degree of semantic information while retaining some detailed features. Based on this characteristic, this embodiment makes targeted lightweight improvements and proposes a bilayer feature fusion module (BFFM), mainly by eliminating the introduction of middle-level feature maps, making BFFM more focused on the bilayer feature fusion of high-level and low-level feature maps, and further optimizing performance. In this step, the high-level image, high-level enhanced feature map, and low-level feature map are input into the bilayer feature fusion module for processing to generate a feature fusion enhanced map; the process of the bilayer feature fusion module processing the input image includes:

[0107] a) Process the low-level feature map and high-level image through reflection convolution to align the space and channels of the feature map output by the reflection convolution with the high-level enhanced feature map:

[0108] F b ' = RefConv(F b )

[0109] F i = RefConv(I u )

[0110] where RefConv is the reflection convolution operation; F b is the low-level feature map, the convolution kernel size k of the reflection convolution operation for processing F b is 5, the convolution kernel stride s is 2, and the reflection padding size p is 2; I u is the high-level image, the convolution kernel size k of the reflection convolution operation for processing I u is 3, the convolution kernel stride s is 1, and the reflection padding size p is 1. Through the foregoing processing, the space and channels of the feature maps F′ b , F i are aligned with the high-level enhanced feature map F u '.

[0111] b) Split the feature maps F′ b , F i and F u ' into N sub-feature maps respectively and add a new dimension T. The process is as follows:

[0112]

[0113] where n represents the decomposition step, n = 1, 2, 3,..., N; c is the feature fusion channel number, and in this embodiment, the channel number c here is 128.

[0114] c) Fuse the sub-feature maps through N fusion sub-modules. The fusion process of each fusion sub-module is as follows:

[0115] First, splice the three sub-feature maps and along the direction of the penultimate dimension to obtain the feature map Subsequently, multiply the three feature sub-maps element-wise with to obtain three cross-attention weights and The element-wise multiplication operation process is as follows:

[0116]

[0117] where ⊙ is the element-wise multiplication operation. Subsequently, splice and sum all the cross-attention weights along the direction of dimension S, and then perform feature normalization along the last dimension to obtain the double-layer attention weight of a single sub-module The calculation process of the attention weight is as follows:

[0118]

[0119] where is splicing along the direction of dimension S.

[0120] d) Fuse the double-layer attention weight with the three sub-feature maps without adding dimension T F i n , in the grouped dimension of the sub-module to obtain a more informative feature map The weighted summation process is as follows:

[0121]

[0122] e) Upsample F bi by 8 times using bilinear interpolation to obtain the enhanced feature map Meanwhile, to effectively reduce the loss of low-level information, perform a skip connection between F′ e and the low-level feature map F b to obtain the feature fusion enhanced map The above process is as follows:

[0123] F′ e = Up(F bi ), sf = 8

[0124] F e = F e '+ F b

[0125] where Up is bilinear interpolation upsampling and sf is the upsampling scale factor.

[0126] 5) Input the original low-light image, high-level enhanced feature map, and feature fusion enhanced map of the fan blade into the image iterative optimization module for processing to generate an enhanced low-light image. The processing of the input image by the iterative optimization module in this step includes:

[0127] 1) Concatenate the feature fusion enhanced map and the high-level enhanced feature map in the channel direction to obtain a mixed feature map Subsequently, use bilinear interpolation to upsample F fu by 8 times to obtain a more informative iterative feature map The feature concatenation and upsampling processes are as follows:

[0128]

[0129] F i = Up(F fu ), sf = 8

[0130] 2) Decompose the iterative feature map to obtain iterative factors, including: Unfolding the iterative feature map F i in the channel direction and decomposing it into 2N iterative factors F n H×W×3 for image iterative optimization. Each iterative factor corresponds to a different feature representation. The feature splitting process is as follows:

[0131]

[0132] 3) Iterative optimization: Apply the iterative factors to the input original low-light image I for e times of iterative optimization to generate an enhanced low-light image I e , and the iterative optimization process is as follows:

[0133]

[0134] When n = 1, I0 is the input original low-light image, and I e is the enhanced low-light image output after iterative optimization.

[0135] Image iterative optimization can effectively reduce the interference and degradation effects of images in low-light and dark environments, especially performing well in processing complex scenes.

[0136] II) Input the enhanced low-light image obtained in step 5) into the target detection model for processing to obtain the fan blade fault detection result. In this embodiment, the target detection model can use the classic YOLOv5l. Of course, other types can also be selected for the target detection model.

[0137] In this embodiment, the Bi-Enhancer model and the object detection model are cascaded, and the output of the Bi-Enhancer model is directly fed into the backbone network of the detector model; the Bi-Enhancer model and the object detection model are trained in an end-to-end manner. During the training process, the Bi-Enhancer model is implicitly optimized through the loss function of the object detection model, and no separate loss function is set. The ExDark dataset can be used for training. The long side of the image is set to 608, the short side ranges from 320 to 512, the image batch size is set to 8, and the training samples can be geometrically transformed by random horizontal flipping and vertical flipping to improve the generalization ability of the model.

[0138] 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 preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the purpose and scope of the technical solutions of the present invention, and they should all be covered by the scope of the claims of the present invention.

Claims

1. A method for detecting faults in fan blades under low-light environments, characterized in that: It includes the following steps: I) Enhance the image of the fan blade collected under low-light conditions through the Bi-Enhancer model. The Bi-Enhancer model includes a reflection convolutional layer, a lightweight feature aggregation module, a high-level feature enhancement module, a double-layer feature fusion module, and an image iterative optimization module. The enhancement process includes: 1) Input the original low-light image of the fan blade into the reflection convolution for processing to generate a high-level image; 2) Input the original low-light image into the lightweight feature aggregation module for processing to generate a low-level feature map; input the high-level image into the lightweight feature aggregation module for processing to generate a high-level feature map; 3) Input the high-level feature map into the high-level feature enhancement module for processing to generate a high-level enhanced feature map; 4) Input the high-level image, the high-level enhanced feature map, and the low-level feature map into the double-layer feature fusion module for processing to generate a feature fusion enhanced map; 5) Input the original low-light image of the fan blade, the high-level enhanced feature map, and the feature fusion enhanced map into the image iterative optimization module for processing to generate an enhanced low-light image; II) Input the enhanced low-light image obtained in step 5) into the target detection model for processing to obtain the fan blade fault detection result.

2. The method for detecting the fault image of the fan blade in a low-light environment according to claim 1, wherein: In step 1), the process of generating a high-level image by the reflection convolution is as follows: I u = RefConv(I) Wherein: RefConv represents a reflection convolution operation; I is the original low-light image of the fan blade, and I u is the generated high-level image.

3. The method for detecting the fault image of the fan blade in a low-light environment according to claim 2, characterized in that: In step 2), the process of processing the input image by the lightweight feature aggregation module includes: F1 = SiLU[Conv(I in )] F3 = SiLU[Conv(F2)] F5 = SiLU[Conv(F4)] F out = SiLU[Conv(F6)] Among them: Conv is a conventional convolution operation, and SiLU is an activation function. is the feature concatenation along the channel direction, I in is the input image, F out is the output image.

4. The method for detecting the fault image of the fan blade in low-light environment according to claim 3, wherein: In step 3), The process of processing the input high-level feature map by the high-level feature enhancement module includes: F s = FuseConv(F u ) F sr = FuseConv(F s ) F d = DWConv(F sr ) F u ' = DWConv(F d ) Among them, F u is the high-level feature map, and F u ' is the high-level enhanced feature map; FuseConv(F u ) = SiLU[BatchNorm(C1onv(F u )) + BatchNorm(C2onv(F u ))] FuseConv(F s ) = SiLU[BatchNorm(C1onv(F s )) + BatchNorm(C2onv(F s ))] DWConv(F sr ) = SiLU[Point_C2onv(Depth_C1onv(F sr ))] DWConv(F d ) = SiLU[Point_C2onv(Depth_C1onv(F d ))] Among them, F u is the high-level feature map, F u ' is the high-level enhanced feature map, BatchNorm is the regularization operation, C1onv and C2onv are conventional convolution operations, Depth_C1onv is the depth convolution operation, Point_C2onv is the pointwise convolution operation, and SiLU is the activation function.

5. The method for detecting the fault image of the fan blade in low light environment according to claim 4, characterized in that: In step 4), the process of processing the input image by the double-layer feature fusion module includes: a) Process the low-level feature map and the high-level image through the reflection convolution to align the space and channels of the feature map output by the reflection convolution with the high-level enhanced feature map: F′ b = RefConv(F b ) F i = RefConv(I u ) Among them, RefConv is the reflection convolution operation, F b is the underlying feature map, and I u is the high-level image; b) Split the feature maps F′ b , F i and F u ' into N sub - feature maps respectively and add a new dimension T. The process is as follows: Where n represents the decomposition step, n = 1, 2, 3,..., N; c is the number of feature fusion channels; c) Fuse the sub-feature maps through N fusion sub-modules. The fusion process of each fusion sub-module is as follows: First, splice the three sub-feature maps and along the direction of the penultimate dimension. H is the height of the original low-light image, W is the width of the original low-light image, and C is the number of channels of the original low-light image, to obtain the feature map Subsequently, multiply the three feature sub-maps with element-wise to obtain three cross-attention weights and The element-wise multiplication operation process is as follows: Among them, ⊙ is the element multiplication operation. Subsequently, all cross-attention weights are concatenated and summed along the direction of dimension S, and then feature normalization is performed along the last dimension to obtain the double-layer attention weights of a single sub-module. The calculation process of the attention weights is as follows: Among them, is spliced along the dimension S direction; d) Combine the double-layer attention weights through weighted summation with the three sub-feature maps without adding dimension T F i n , and perform fusion on the grouped dimension of the sub-module to obtain a more informative feature map The weighted summation process is as follows: e) Upsample F by 8 times using bilinear interpolation method bi to obtain an enhanced feature map Meanwhile, in order to effectively reduce the loss of underlying information, F e ' is skip-connected with the underlying feature map F b to obtain a feature fusion enhanced map The above process is as follows: F e ' = Up(F bi ), sf = 8 F e = F e '+ F b Where Up is bilinear interpolation upsampling and sf is the upsampling scale factor.

6. The method for detecting the fault image of the fan blade in a low-light environment according to claim 5, characterized in that: In step 5), the process of processing the input image by the iterative optimization module includes: 1) Concatenate the feature fusion enhanced map and the high-level enhanced feature map in the channel direction to obtain a mixed feature map Subsequently, use bilinear interpolation to upsample F fu by 8 times to obtain a more informative iterative feature map The feature concatenation and upsampling processes are as follows: F i = Up(F fu ), sf = 8 2) Decomposing the iterative feature map to obtain iterative factors includes: Unfolding the iterative feature map F i in the channel direction and decomposing it into 2N iterative factors F for image iterative optimization n H×W×3 , each iterative factor corresponding to a different feature representation. The feature splitting process is as follows: 3) Iterative optimization: Apply the iterative factor to the input original low-light image I for e times of iterative optimization to generate the enhanced low-light image I e , and the iterative optimization process is as follows: When n = 1, I0 is the original low-light image input, and I e is the enhanced low-light image output after iterative optimization.

7. The method for detecting the fault image of the fan blade in a low-light environment according to claim 1, wherein: The Bi-Enhancer model and the target detection model are cascaded, and the output of the Bi-Enhancer model is directly passed into the backbone network of the detector model; During the training of the Bi-Enhancer model and the target detection model, the Bi-Enhancer model is implicitly optimized through the loss function of the target detection model, and no separate loss function is set.

Images