Brightness-guided dual-domain reconstruction quadruped robot with low-light target detection method

Abstract

The application discloses a low-light target detection method for a luminance-guided dual-domain reconstruction quadruped robot, and belongs to the field of computer vision processing calculation. The method comprises the following steps: constructing a prior-guided dual-domain shallow feature reconstruction module, and replacing the first Conv module at the input end of the YOLOv11 backbone network with the module, keeping the remaining network structure and parameters of the YOLOv11 network unchanged. The module generates a local variance map prior-guided illumination-invariant feature and a luminance-prior-enhanced color compensation feature in parallel, and the two features are fused, convolved, batch-normalized and SiLU-mapped to form a first layer of reconstructed features consistent with the interface of the original first Conv module, and then the subsequent network completes the detection. The method can be used for low-light target detection of a quadruped robot in underground space, night patrol and complex unstructured environment, and can improve the detection stability under weak texture, small target and complex lighting conditions.

Description

Technical Field

[0001] This invention belongs to the field of computer vision processing technology, specifically relating to a low-light target detection method for a brightness-guided dual-domain reconstruction quadrupedal omnipotent intelligent robot. Background Technology

[0002] Quadrupedal anthropomorphic intelligent robots possess strong obstacle-crossing capabilities and mobility in unstructured terrain, and are gradually being applied in coal mine inspections, underground space exploration, nighttime security, post-disaster search and rescue, and remote operations in high-risk areas. In these scenarios, visual sensors are typically affected simultaneously by low illumination, localized strong light, light source flicker, motion blur, sensor noise, and reflected glare. Within the same image, there may be both excessively dark and excessively bright areas. This type of compound degradation not only reduces the contrast between the target and the background but also causes color shift, edge fragmentation, and loss of texture for small targets, leading to missed detections, false detections, and bounding box offsets in the detection network.

[0003] Existing technologies often employ a sequential "enhancement-then-detection" approach, where Retinex decomposition, histogram adjustment, or depth enhancement networks are first used to improve image visibility before the enhancement result is input into the target detector. This approach prioritizes human visual quality as its primary optimization objective, and the bright areas, color shifts, and pseudo-textures generated during the enhancement stage may be mistakenly identified as target features by subsequent detectors. Furthermore, the lack of task constraints between the enhancement and detection networks, coupled with inconsistencies in their loss functions, data distributions, and optimization objectives, can easily lead to visually brighter enhancement results but decreased detection performance.

[0004] Another type of approach incorporates attention, feature pyramids, or illumination-aware units into the middle or neck layers of the detection network. While these methods can recalibrate already formed deep semantic features, low-light images first undergo convolution and downsampling in the initial layer, potentially compressing or mixing weak edges, color differences, and dark textures. Once shallow information suffers irreversible loss during downsampling, subsequent deep enhancements struggle to recover the original details. Therefore, simply adding feature enhancement units in the middle and later stages cannot fundamentally address the representational loss caused by low-light degradation at the network's entry point. Existing illumination-invariant modeling typically relies on fixed-scale local statistics, single convolutional kernels, or simple channel normalization to generate invariant features. Fixed scales struggle to simultaneously adapt to large smooth dark areas, fine target edges, and high-frequency textures; relying solely on proportional or normalized features easily erases target contours and local intensity differences. On the other hand, color and brightness are highly coupled in the RGB space, and direct convolution will encode both brightness fluctuations and color residuals. Although existing color space transformations can separate chromaticity and intensity, if the chromaticity branch lacks brightness reliability constraints, color noise in extremely dark areas may still be excessively amplified, which greatly limits the further improvement of low-light target detection performance. Summary of the Invention

[0005] The purpose of this invention is to provide a low-light target detection method for a brightness-guided dual-domain reconstruction quadrupedal anthropomorphic intelligent robot, in order to solve the problems in the prior art such as inconsistency between enhancement and detection targets, premature loss of shallow weak features, difficulty in balancing illumination-invariant representation and edge details, easy amplification of color noise in extremely dark areas, and inability to share multi-branch priors.

[0006] The technical solution adopted in this invention is a low-light target detection method for a brightness-guided dual-domain reconstruction quadrupedal anthropomorphic intelligent robot, which specifically includes the following steps: Step 1: Construct a priori-guided dual-domain shallow feature reconstruction module and replace the first Conv module at the input of the YOLOv11 backbone network with this module, while retaining the rest of the original YOLOv11 backbone, neck and detection head structures to build a low-light target detection model; Step 2: Collect and organize real low-light scene images, complete annotation and size standardization, and establish a low-light target detection dataset; Step 3: Input the images from the low-light target detection dataset into the low-light target detection model, and design a combined loss function to complete the model optimization training; Step 4: Input the low-light image to be detected into the optimized and trained low-light target detection model, and output the target category, confidence score and location results.

[0007] The invention is further characterized by: The prior-guided dual-domain shallow feature reconstruction module includes a shallow feature generation unit and a dual-domain feature reconstruction unit. The shallow feature generation unit includes a parallel local variance prior-guided illumination-invariant feature generation module and a color decoupling and compensation feature generation module.

[0008] Step 2 specifically includes: Real-world low-light scene images are collected, and the raw image data is cleaned and filtered to remove blurry, overexposed, and invalid samples without targets. The image resolution is standardized to 1280×720 pixels. Data regularization and standardization are then completed to obtain a low-light target detection dataset that can be used for model training and testing.

[0009] Step 3 specifically includes the following sub-steps: Step 3.1: Input the images from the low-light target detection dataset into the shallow feature generation unit. The illumination-invariant feature generation module and the color decoupling and compensation feature generation module, guided by local variance prior, process the input images simultaneously. The illumination-invariant feature generation module, guided by local variance prior, outputs illumination-invariant features and a brightness map. The brightness map serves as the brightness prior for the color decoupling and compensation feature generation module. The color decoupling and compensation feature generation module combines the brightness prior to output color compensation features enhanced by the brightness prior. Step 3.2: Input the illumination-invariant features and the brightness-prior-enhanced color compensation features into the dual-domain feature reconstruction unit to generate the first-layer reconstructed features with the same output size and number of channels as the original first Conv module of YOLOv11; Step 3.3: Input the reconstructed features from the first layer into the subsequent YOLOv11 network, and complete the model optimization training by combining loss functions to obtain the trained low-light target detection model.

[0010] The illumination-invariant feature generation module guided by local variance prior includes a parallel candidate feature generation network and a local variance prior guiding network; it also includes a gated residual fusion module. Candidate feature generation network is used to extract attention-enhanced candidate features. The initial illumination invariant characteristics are obtained through inter-channel division. ; Local variance prior guided network is used to compute local variance map prior from input image. ; The gated residual fusion module will use the local variance map as a priori. As weights for attention enhancement candidate features Invariant features with initial illumination Adaptive fusion is performed to obtain illumination-invariant features. , means as follows: .

[0011] The local variance prior-guided network processes the input image as follows: First, the input image The R, G, and B channels are used to obtain the maximum value at each pixel location to get a single-channel luminance map. , means as follows:

[0012] In the formula, and These represent the horizontal and vertical coordinates of the image pixel positions, respectively. Indicates in the channel c Take the maximum value from the middle. , These represent the height and width of the input image, respectively. Indicates the input image in channels c Pixel value at; Meanwhile, brightness map As a priori output of brightness; Then, the single-channel brightness map Intra-channel feature mapping is performed using a 5×5 depthwise convolution kernel to obtain the brightness feature map. , is represented as:

[0013] In the formula, Indicates a depthwise convolution operation; Next, a depthwise convolution with a kernel size of 5×5 is used to aggregate neighborhood information within each channel, and the ith value is calculated. Each channel is located at coordinate position The local mean at point is expressed as:

[0014] In the formula, Indicates the first Each channel is located at coordinate position Local mean at that location Indicates The 5×5 neighborhood centered on the center Represents brightness feature mapping In the Coordinate position in each channel The characteristics of the place, u and v This represents the horizontal and vertical offsets within the local convolution window; Based on this, brightness feature mapping In the Coordinate position in each channel Features of the place With the Each channel is located at coordinate position Local mean at The squared deviations are then aggregated using a 5×5 depthwise convolution to obtain the coordinate positions. Prior local variance plot , means as follows: .

[0015] The candidate feature generation network processes the input image as follows: First, the input image The results are processed by 3×3 convolutional layers and 5×5 convolutional layers respectively to obtain local responses at different scales. The two results are then merged along the channel direction to obtain multi-scale candidate features. Secondly, multi-scale candidate features are input into the channel-space joint attention mechanism: first, global average pooling, global max pooling, and shared fully connected mapping are used to obtain channel weights; then, average pooling, max pooling, and 7×7 convolutions are performed on the channel-recalibrated features to obtain spatial weights; finally, attention-enhanced candidate features are output. ; Subsequently, illumination-invariant modeling was performed on the attention-enhancing candidate features. The impact of global illumination scale variations was mitigated through numerical stabilization and adjacent channel ratio division. Specifically, this included: To avoid zero denominators in channel ratio division, attention-enhanced candidate features are used. The numerical stabilization process is performed beforehand, as shown below:

[0016]

[0017] In the formula, It is a constant. For symbolic functions, To enhance attention to candidate features The processed value; Attention-enhanced candidate features after function processing; By performing proportional division on adjacent channels in a cyclic manner and then stitching them together along the channel dimension, the initial illumination-invariant feature is obtained, as shown below:

[0018] In the formula, Indicates the first i Feature map of each channel This indicates a splicing operation at the channel level. This indicates that the initial illumination remains unchanged.

[0019] The color decoupling and compensation feature generation module includes an HVI color decoupling unit, a color compensation feature extraction unit, and a brightness prior-guided enhancement unit connected in sequence. The image processing procedure of the color decoupling and compensation feature generation module on the input low-light target detection dataset is as follows: First, the input image is a standard three-channel RGB image. Input to HVI color decoupling unit: Calculate the input image brightness value saturation Hue Let the maximum and minimum channel responses at the same pixel be respectively... T and m Then saturation Calculate using the following formula:

[0020]

[0021]

[0022] In the formula, It is a constant; Indicates in the channel c Take the minimum value from the middle. Introducing a color sensitivity adjustment factor joint saturation Hue Polarization mapping is performed to obtain hue components. and vertical components Define intensity components I And assign the brightness value to the intensity component. I This constitutes the HVI decoupling representation, as shown below:

[0023] In the formula, Represents the hue value of a single pixel; Among them, color sensitivity adjustment factor Calculate using the following formula:

[0024] In the formula, This indicates the color sensitivity adjustment parameter; Secondly, the color compensation feature extraction unit encodes the hue component and the vertical component: , Color compensation features are obtained after splicing and chromaticity residual coding. At the same time, the intensity component I Strength constraint features are obtained through strength constraint encoding. , means as follows:

[0025]

[0026] In the formula, , They represent and Convolution operation; where the chroma residual encoding consists of 3×3 convolutions and the intensity constraint encoding consists of 5×5 convolutions; Subsequently, the luminance prior-guided enhancement unit analyzed the luminance map. Color compensation features and strength constraint features Joint fusion is performed: After spatial size matching of the luminance map with the color compensation features and intensity constraint features, they are stitched together along the channel dimension to obtain the color compensation features with luminance prior enhancement. , means as follows:

[0027] In the formula, This indicates a splicing operation at the channel dimension.

[0028] The dual-domain feature reconstruction unit includes channel splicing units connected in series. Convolutional layers, batch normalization layers, and the SiLU activation function; The output illumination-invariant feature Color compensation features with prior brightness enhancement Input a dual-domain feature reconstruction unit and output a first-layer reconstructed feature that combines illumination robustness and color perception capability. , means as follows:

[0029] In the formula, This indicates a channel splicing operation. This indicates a batch normalization operation. This represents the SiLU activation function. express Convolution operation.

[0030] The combined loss function consists of bounding box regression loss. Targeted loss and classification loss The weighted composition is represented as follows:

[0031] In the formula, , , The bounding box regression loss is respectively Targeted loss and classification loss The weight, Denotes the combined loss function; where, , , All numbers are positive and satisfy the following conditions: ; Bounding box regression loss , means as follows:

[0032] In the formula, This represents an improved IoU loss calculated based on the overlap between the predicted bounding box and the ground truth bounding box. This represents the distributed regression loss; The internal adjustment coefficient of distributed regression loss in bounding box regression loss; Targeted loss , means as follows:

[0033] In the formula, The target score predicted by the model. For soft labels related to IoU quality; Indicates loss of quality focus; Based on the type of the detection head, the output shape is similar, and the classification loss is calculated. Use either multi-class cross-entropy or class-by-class binary cross-entropy summation: Using a multi-class cross-entropy approach: Samples are labeled with hot tags, with the target class label set to 1 and other class labels set to 0, as shown below:

[0034] In the formula, Indicates the total number of categories. Indicate category The true label, To output the category to which the sample belongs The predicted probability; The classification loss is obtained by summing the binary cross-entropy values ​​for each category: the binary cross-entropy is calculated separately for each category and then summed to obtain the total classification loss, as shown below: .

[0035] The beneficial effects of this invention are: This invention presents a brightness-guided dual-domain reconstruction method for low-light target detection in quadrupedal anthropomorphic intelligent robots. Compared with existing technologies, this method utilizes a local variance prior-guided illumination-invariant feature generation module to unify the modeling of multi-scale candidate features, channel-space joint attention, channel-ratio illumination-invariant modeling, and local variance map prior-guided gated residual fusion. This enables the network to suppress illumination changes in smooth dark areas and retain effective structure in edge and texture regions. The color decoupling and compensation feature generation module separates chromaticity and intensity information in the HVI space and uses the luminance map as a luminance prior to guide and enhance color compensation features, mitigating color shift and detail loss in low-light scenes. Furthermore, a dual-domain feature reconstruction unit maps two types of shallow features to network inputs consistent with the first Conv module, thereby improving detection robustness in complex low-light environments without altering the subsequent backbone, neck, and detection head structure of YOLOv11. The method of this invention has strong generalization ability in complex low-light, low-visibility and weak feature scenes, providing effective technical support for quadrupedal intelligent robots to perform real-time visual perception and target detection tasks in underground spaces, nighttime inspections and complex unstructured environments. Attached Figure Description

[0036] Figure 1 This is a network architecture diagram of the low-light target detection method for a brightness-guided dual-domain reconstruction quadrupedal omnipotent intelligent robot according to the present invention; Figure 2 This is a schematic diagram of the structure of the illumination-invariant feature generation module guided by local variance prior in the low-light target detection method for a quadrupedal omnipotent intelligent robot under the brightness-guided dual-domain reconstruction method of the present invention. Figure 3 This is a schematic diagram of the color decoupling and compensation feature generation module in the brightness-guided dual-domain reconstruction quadrupedal omnipotent intelligent robot low-light target detection method of the present invention; Figure 4 This is a comparison image of the original image and the local variance map in the brightness-guided dual-domain reconstruction quadrupedal omnipotent intelligent robot low-light target detection method of the present invention; Figure 5 This is a visualization of the RGB image to HVI image conversion in the brightness-guided dual-domain reconstruction quadrupedal omnipotent intelligent robot low-light target detection method of the present invention; Figure 6 This is a comparison chart of feature extraction results from different modules in simulation experiment 1 of this invention; Figure 7 These are visual effect diagrams of different module combinations in simulation experiment 1 of this invention; Figure 8 This is a comparison chart of the detection results of simulation experiment 2 of this invention on the Exdark dataset; Figure 9 This is a comparison chart of the detection results of simulation experiment 3 of this invention on the RTTS dataset. Detailed Implementation

[0037] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0038] Example 1 When quadrupedal anthropomorphic intelligent robots perform underground inspections, nighttime operations, and unstructured scene detection tasks, the acquired images are easily affected by factors such as insufficient lighting, small target scale, frequent occlusion, and complex background textures. Therefore, the detection network is required to have high detection stability, operating efficiency, and multi-scale feature representation capabilities. YOLOv11 has a good accuracy-speed balance and can adapt to the real-time vision processing requirements of the robot, so it is used as the baseline architecture of this invention.

[0039] Combination Figure 1 The present invention provides a low-light target detection method for a brightness-guided dual-domain reconstruction quadrupedal omnipotent intelligent robot, which specifically includes the following steps: Step 1: Construct a priori-guided dual-domain shallow feature reconstruction module and replace the first Conv module at the input of the YOLOv11 backbone network with this module, while retaining the rest of the original YOLOv11 backbone, neck and detection head structures to build a low-light target detection model; The first Conv module at the input of the YOLOv11 backbone network is a basic convolutional unit consisting of a two-dimensional convolutional layer, a batch normalization layer, and a SiLU activation function stacked sequentially. The prior-guided dual-domain shallow feature reconstruction module includes a shallow feature generation unit and a dual-domain feature reconstruction unit. Combined with... Figure 2 and Figure 3 As shown, the shallow feature generation unit includes a parallel locally variance-guided illumination-invariant feature generation module and a color decoupling and compensation feature generation module. The brightness map generated by the locally variance-guided illumination-invariant feature generation module serves as the brightness prior for the color decoupling and compensation feature generation module. The output channel number and feature resolution of the prior-guided dual-domain shallow reconstruction module match the input requirements of subsequent layers in the YOLOv11 backbone network, enabling the reconstructed features to be directly fed into the backbone network for step-by-step feature extraction.

[0040] Step 2: Collect and organize real low-light scene images, complete annotation and size standardization, and establish a low-light target detection dataset; Step 3: Input the images from the low-light target detection dataset into the low-light target detection model, and design a combined loss function to complete the model optimization training; Step 4: Input the low-light image to be detected into the optimized and trained low-light target detection model, and output the target category, confidence score and location results.

[0041] Example 2 Step 2 specifically includes: acquiring scene images in real low-light environments; cleaning and filtering the acquired raw image data to remove blurry, overexposed, and targetless invalid samples; standardizing the image resolution to 1280×720 pixels; and completing data regularization and standardization to obtain a low-light target detection dataset that can be used for model training and testing. Before training and testing, in this embodiment, the low-light target detection dataset is divided into a training set and a test set in an 8:2 ratio.

[0042] Example 3 Based on the established low-light target detection model, the low-light target detection dataset is input into the model, and a combined loss function is designed to complete the model optimization training. The specific steps include the following: Step 3.1: Input the images from the low-light target detection dataset into the shallow feature generation unit. The illumination-invariant feature generation module and the color decoupling and compensation feature generation module, guided by local variance prior, process the input images simultaneously. The illumination-invariant feature generation module, guided by local variance prior, outputs illumination-invariant features and a brightness map. The brightness map serves as the brightness prior for the color decoupling and compensation feature generation module. The color decoupling and compensation feature generation module combines the brightness prior to output color compensation features enhanced by the brightness prior. Step 3.2: Input the illumination-invariant features and the brightness-prior-enhanced color compensation features into the dual-domain feature reconstruction unit to generate the first-layer reconstructed features with the same output size and number of channels as the original first Conv module of YOLOv11; Step 3.3: Input the reconstructed features from the first layer into the subsequent YOLOv11 network, and complete the model optimization training by combining loss functions to obtain the trained low-light target detection model.

[0043] Example 4 Based on Example 3, in step 3.1, the illumination-invariant feature generation module guided by local variance prior in the shallow feature generation unit includes a parallel candidate feature generation network and a local variance prior guiding network; it also includes a gated residual fusion module. Candidate feature generation network is used to extract attention-enhanced candidate features. The initial illumination invariant characteristics are obtained through inter-channel division. ; Local variance prior guided network is used to compute local variance map prior from input image. ; The gated residual fusion module will use the local variance map as a priori. As weights for attention enhancement candidate features Invariant features with initial illumination Adaptive fusion is performed to obtain illumination-invariant features. , means as follows: .

[0044] Specifically, in combination Figure 2 The local variance prior-guided network processes the input image as follows: First, the input image The R, G, and B channels are taken at their maximum values ​​at each pixel location to obtain a single-channel luminance map. , means as follows:

[0045] In the formula, and These represent the horizontal and vertical coordinates of the image pixel positions, respectively. Indicates in the channel c Take the maximum value from the middle. , These represent the height and width of the input image, respectively. Indicates the input image in channels c Pixel value at; Meanwhile, brightness map As a priori output of brightness; Then, the brightness map Intra-channel feature mapping is performed using a 5×5 depthwise convolution kernel to obtain the brightness feature map. , is represented as:

[0046] In the formula, Indicates a depthwise convolution operation; Edges and textured regions in real low-light images are often accompanied by significant brightness abrupt changes, making them difficult to describe using a uniform local smoothing model. Therefore, this invention constructs a local variance map to characterize the degree of gray-level fluctuation at different locations and uses it as prior information for feature adjustment, enabling the network to employ differentiated feature extraction methods for smooth and edge regions. Specifically, A depthwise convolution with a kernel size of 5×5 is used to aggregate neighborhood information within each channel, and the i-th... Each channel is located at coordinate position The local mean at point is expressed as:

[0047] In the formula, Indicates the first Each channel is located at coordinate position Local mean at that location Indicates The 5×5 neighborhood centered on the center Represents brightness feature mapping In the Coordinate position in each channel The characteristics of the place, u and v This represents the horizontal and vertical offsets within the local convolution window; Based on this, brightness feature mapping In the Coordinate position in each channel Features of the place With the Each channel is located at coordinate position Local mean at The squared deviations are then aggregated using a 5×5 depthwise convolution to obtain the coordinate positions. Prior local variance plot , means as follows: .

[0048] Depend on Figure 4As can be seen, the target contour, densely textured regions, and locations of abrupt brightness changes exhibit high prior responses, while the responses in smooth background regions are relatively low. Therefore, the local variance map can reflect the degree of local variation at different spatial locations, providing guidance for subsequent networks to selectively preserve edge details, enhance texture information, and adjust features in smooth regions.

[0049] Specifically, the candidate feature generation network processes the input image as follows: First, the input image The results are processed by 3×3 convolutional layers and 5×5 convolutional layers respectively to obtain local responses at different scales. The two results are then merged along the channel direction to obtain multi-scale candidate features. Secondly, multi-scale candidate features are input into the channel-space joint attention mechanism: first, global average pooling, global max pooling, and shared fully connected mapping are used to obtain channel weights; then, average pooling, max pooling, and 7×7 convolutions are performed on the channel-recalibrated features to obtain spatial weights; finally, attention-enhanced candidate features are output. ; Subsequently, illumination-invariant modeling was performed on the attention-enhancing candidate features. The impact of global illumination scale variations was mitigated through numerical stabilization and adjacent channel ratio division. Specifically, this included: To avoid zero denominators in channel ratio division, attention-enhanced candidate features are used. The numerical stabilization process is performed beforehand, as shown below:

[0050]

[0051] In the formula, It is a constant. For symbolic functions, To enhance attention to candidate features The processed value; Attention-enhanced candidate features after function processing; By performing proportional division on adjacent channels in a cyclic manner and then stitching them together along the channel dimension, the initial illumination-invariant feature is obtained, as shown below:

[0052] In the formula, Indicates the first i Attention-enhanced candidate features for each channel after processing by a function. This indicates a splicing operation at the channel level. This indicates that the initial illumination remains unchanged.

[0053] The gated residual fusion module will use the local variance map as a priori. As weights for attention enhancement candidate features Invariant features with initial illumination Adaptive fusion is performed to obtain illumination-invariant features. , means as follows: .

[0054] As can be seen from the above formula, the weights are applied in the smooth regions of the image. The size is relatively large, therefore the final illumination remains unchanged. More dependent on the invariance of initial illumination In uneven or edge areas, The weighting increases, making Prefer to use attention-enhanced candidate features Through this adaptive feature fusion method, the model can effectively extract illumination-invariant features, thereby maintaining more stable and robust performance under different lighting conditions.

[0055] The color decoupling and compensation feature generation module in the shallow feature generation unit includes an HVI color decoupling unit, a color compensation feature extraction unit, and a luminance prior-guided enhancement unit connected in sequence; combined with Figure 3 The color decoupling and compensation feature generation module processes images from the input low-light target detection dataset as follows: First, the input image is a standard three-channel RGB image. Input to HVI color decoupling unit: Calculate the input image brightness value saturation Hue Let the maximum and minimum channel responses at the same pixel be respectively... T and m Then saturation Calculate using the following formula:

[0056]

[0057]

[0058] In the formula, It is a constant; Indicates in the channel c Take the minimum value from the middle. Introducing a color sensitivity adjustment factor joint saturation Hue Polarization mapping is performed to obtain hue components. and vertical components Define intensity components I And assign the brightness value to the intensity component. I This constitutes the HVI decoupling representation, as shown below:

[0059] In the formula, Represents the hue value of a single pixel; Among them, color sensitivity adjustment factor Calculate using the following formula:

[0060] In the formula, This indicates the color sensitivity adjustment parameter; Secondly, the color compensation feature extraction unit encodes the hue component and the vertical component: , Color compensation features are obtained after splicing and chromaticity residual coding. At the same time, the intensity component I Strength constraint features are obtained through strength constraint encoding. , means as follows:

[0061]

[0062] In the formula, , They represent and Convolution operation; where the chroma residual encoding consists of 3×3 convolutions and the intensity constraint encoding consists of 5×5 convolutions; Subsequently, the luminance prior-guided enhancement unit analyzed the luminance map. Color compensation features and strength constraint features Joint fusion is performed: After spatial size matching of the luminance map with the color compensation features and intensity constraint features, they are stitched together along the channel dimension to obtain the color compensation features with luminance prior enhancement. , means as follows:

[0063] In the formula, This indicates a splicing operation at the channel dimension.

[0064] like Figure 5As shown, the color decoupling and compensation feature generation module enhances the color compensation capability in low-light images. In summary, the color decoupling and compensation feature generation module fully utilizes the advantages of the HVI color space, improving the model's understanding of the image's color structure, thus more effectively enhancing the overall image quality when processing images under complex lighting conditions.

[0065] Example 5 Based on Example 4, in step 3.2, the dual-domain feature reconstruction unit includes a channel splicing unit connected in series. Convolutional layers, batch normalization layers, and the SiLU activation function; The output illumination-invariant feature Color compensation features with prior brightness enhancement Input a dual-domain feature reconstruction unit and output a first-layer reconstructed feature that combines illumination robustness and color perception capability. , means as follows:

[0066] In the formula, This indicates a channel splicing operation. This indicates a batch normalization operation. This represents the SiLU activation function. express Convolution operation.

[0067] Example 6 Based on Example 5, the combined loss function is derived from the bounding box regression loss. Targeted loss and classification loss A linear combination based on preset weights is represented as follows:

[0068] In the formula, , , These represent the bounding box regression loss, respectively. Targeted loss and classification loss The corresponding weighting coefficients, Let represent the combined loss function; all three weight coefficients are positive and satisfy:

[0069] Among them, the bounding box regression loss weights The value range is 0.45 to 0.60, which is the target loss weight. The value range is 0.20 to 0.35, which is the classification loss weight. The value range is 0.15 to 0.30. In this embodiment, the weights of the bounding box regression loss, the targetability loss, and the classification loss are 0.50, 0.30, and 0.20, respectively, that is, the weight ratio of the three is 5:3:2.

[0070] The weights are determined through a grid search on the validation set: within the range of values, three candidate combinations with a weight sum of 1 are constructed using a step size of 0.05; the model is trained separately under the same conditions of training data, training epochs, and other hyperparameters, and the combination with the highest mAP50:95 on the validation set is selected as the final value. When multiple combinations have the same mAP50:95, the combination with higher recall is selected first. Once the weights are determined, they remain fixed during a single model training iteration.

[0071] Among them, bounding box regression loss It measures the difference between the predicted bounding box and the true bounding box in terms of location, size, and shape, and consists of an improved IoU loss based on overlap and a distributional regression loss, as shown below:

[0072] In the formula, This represents an improved IoU loss calculated based on the overlap between the predicted bounding box and the ground truth bounding box. This represents the distributed regression loss. This represents the internal adjustment coefficient of the distribution regression loss in the bounding box regression loss, with a value ranging from 0.10 to 0.50. In this embodiment, 0.20 is preferred. The two components of the bounding box regression loss together improve the accuracy of the predicted box in terms of center point location, aspect ratio, and boundary granularity.

[0073] Targeted loss The measure of the probability that a predicted location contains a target is represented as follows:

[0074] In the formula, The target score predicted by the model. For soft labels related to IoU quality; This represents the quality focus loss. The contribution of the goal-orientation loss to the combined loss function is determined by the weighting coefficients. control, The value of is in the range of 0.20 to 0.35, and in this embodiment, 0.30 is preferred. This loss allows the model to give higher confidence to high-quality matching boxes and lower confidence to low-quality or background boxes, thereby improving detection robustness.

[0075] Classification loss It is used to measure the difference between the predicted target category and the true target category. Depending on the category output format of the detection head, the classification loss adopts either the multi-class cross-entropy form or the class-by-class binary cross-entropy accumulation form.

[0076] Using a multi-class cross-entropy approach: Samples are labeled with hot tags, with the target class label set to 1 and other class labels set to 0, as shown below:

[0077] In the formula, Indicates the total number of categories. Indicate category The true label, To output the category to which the sample belongs The predicted probability; The classification loss is obtained by summing the binary cross-entropy values ​​for each category: the binary cross-entropy is calculated separately for each category and then summed to obtain the total classification loss, as shown below: .

[0078] The contribution of classification loss to the combined loss function is determined by the weighting coefficients. control, The value range is 0.15 to 0.30, and in this embodiment, 0.20 is preferred. Classification loss This ensures that the model correctly classifies boxes containing targets, thus improving the accuracy of target classification.

[0079] Example 7 To comprehensively verify the detection capabilities of this invention under low-light conditions, experiments were conducted using two publicly available datasets: ExDark and DarkFace. ExDark covers various complex scenarios, including nighttime roads, indoor low light, localized high brightness, and occlusion interference. These scenarios closely resemble the uneven lighting, shadow variations, and weakened target textures encountered by quadrupedal inspection robots in underground spaces and nighttime operations, making it suitable for examining the model's stability and adaptability in complex low-light environments. DarkFace primarily focuses on nighttime face detection, and its samples generally exhibit low contrast, high noise, and uneven lighting distribution, making it suitable for further analysis of the model's recognition performance on targets with weak features. To examine the model's transferability in other visually degraded scenarios, extended tests were also conducted on the RTTS foggy day dataset. Fog and low light can both reduce image contrast, attenuate texture information, and weaken target features; therefore, RTTS can be used to verify the method's cross-scene generalization performance.

[0080] Experimental environment: The experiments were conducted on an Ubuntu 24.04 operating system, based on the YOLOv11 object detection network, using an NVIDIA RTX 4090 graphics card. The software environment employed the PyTorch framework and Python 3.12. The main parameters for model training are shown in Table 1. The experiments used mAP50, mAP50:95, and Recall as evaluation metrics to measure the model's detection accuracy, overall localization ability under different cross-union ratio (CUP) thresholds, and object detection capability, respectively.

[0081] Table 1 Parameter settings for model training

[0082] Experiment 1: Ablation Experiment.

[0083] To analyze the improvement effect of the low-light target detection model built by the method of this invention under low-light conditions and to clarify the impact of each functional branch on detection performance, a module ablation experiment was conducted using YOLOv11 as the benchmark model. The experiment configured a local variance prior-guided illumination-invariant feature generation module, a color decoupling and compensation feature generation module, and a shallow feature generation unit and a dual-domain feature reconstruction unit composed of both. The actual effects of different configurations were compared from the perspectives of detection metrics and feature response.

[0084] Representative low-light images were selected from the ExDark and DarkFace datasets, and the shallow features generated by different modules were visualized and analyzed. The results are as follows: Figure 6 As shown, raw low-light images generally suffer from insufficient overall illumination, weakened target contours, missing local textures, and background response interference. By introducing a local variance prior-guided illumination-invariant feature generation module, the local variance map prior can focus on guiding edge and texture variation areas. Simultaneously, multi-scale candidate features and channel ratio modeling mitigate the impact of illumination changes, resulting in more stable preservation of target structure and contour information. Using a color decoupling and compensation feature generation module, chromaticity and intensity information in the input image are separated and modeled in the HVI space. Color compensation and luminance prior fusion improve the color attenuation problem in dark areas, making color differences, edge responses, and local textures in the target region more pronounced. Further employing a complete shallow feature generation unit and a dual-domain feature reconstruction unit, the illumination-invariant features and the luminance-prior-enhanced color compensation features achieve complementary expression after dual-domain feature reconstruction. This suppresses non-uniform illumination interference while preserving color and structural information, thus forming shallow features more suitable for subsequent detection network processing.

[0085] Further quantitative comparisons were conducted on the ExDark dataset, and the experimental results are shown in Table 2. The mAP50, mAP50:95, and Recall of the YOLOv11 baseline model were 74.1%, 50.7%, and 64.5%, respectively. When Model 2 was configured only with the illumination-invariant feature generation module guided by local variance prior, the three metrics were 73.8%, 50.6%, and 63.7%, respectively. This result indicates that while enhancing illumination stability alone can improve some structural responses, its detection advantages are not fully realized without color information compensation. When Model 3 was configured only with the color decoupling and compensation feature generation module, the mAP50 and mAP50:95 were 73.9% and 50.3%, respectively, and the Recall reached 65.8%, an improvement of 1.3 percentage points compared to the baseline model. This shows that chromaticity-intensity decoupling and color compensation are beneficial for improving the response intensity and detection probability of weak feature targets, but the improvement in target localization accuracy by a single color branch is still limited. Model 4 is configured with both a local variance prior-guided illumination-invariant feature generation module and a color decoupling and compensation feature generation module. Through dual-domain feature reconstruction, a complete front-end structure is formed. Its mAP50 and mAP50:95 reach 75.1% and 51.1%, respectively, improving upon the baseline model by 1.0 and 0.4 percentage points, achieving the highest detection accuracy among all configurations. The recall is 64.5%, consistent with the baseline model. These results demonstrate that illumination-invariant structural information and color compensation information have a significant complementary relationship. After reconstruction by the dual-domain feature reconstruction unit, they can more effectively improve the shallow feature representation and overall detection performance of YOLOv11 in complex low-light scenes.

[0086] Table 2 Ablation Experiment Results

[0087] To visually examine the detection performance of each functional module in real low-light images, this invention conducts qualitative comparative experiments on the ExDark dataset. The results are as follows: Figure 7 As shown in the figure, the detection results are presented using only the color decoupling and compensation feature generation module, only the illumination-invariant feature generation module guided by local variance prior, and both the complete shallow feature generation unit and the dual-domain feature reconstruction unit. Figure 7 As shown in (a), in extremely low-illuminance scenarios, both individual modules can improve target response, but differences remain in detection box coverage, target localization, and confidence. By employing a complete shallow feature generation unit and a dual-domain feature reconstruction unit module, the model can simultaneously utilize illumination-invariant structural information and color compensation information, resulting in a higher degree of matching between the detection box and the target region, and achieving a more stable detection confidence. Figure 7In the complex nighttime scenes shown in (b) and (c), when using only a single feature generation module, the model is easily affected by background lighting, local occlusion, and changes in target scale, potentially leading to phenomena such as detection box shifts, repetitive responses, or insufficient features for small targets. Dual-domain feature reconstruction can enhance the complementarity between the two types of shallow information, making the target response more focused and the detection box position more accurate. Figure 7 The localized strong light interference scene in (d) and Figure 7 In the distant, small-scale, and low-contrast target scenes in (e), the complete module also demonstrates better adaptability. The local variance map prior can enhance the edge and texture variation areas, and the color decoupling and compensation process can reduce the color decay in dark areas. The combination of the two helps to maintain the target outline and improve the recognizability of weak feature targets, thereby reducing missed detections, false detections, and localization errors.

[0088] In summary, the illumination-invariant feature generation module guided by local variance prior is mainly used to improve the network's adaptability to non-uniform illumination, edge structure, and texture information; the color decoupling and compensation feature generation module is mainly used to improve color representation and weak target response under low-light conditions. The complete shallow feature generation unit and dual-domain feature reconstruction unit, by fusing two types of complementary features, enable the network to obtain more stable target detection results in complex low-light scenes. Ablation experiments thus verify the role of each component module and the effectiveness of their synergistic use.

[0089] Experiment 2: Comparative Experiment.

[0090] This experiment compares the proposed method with various object detection models on the ExDark and DarkFace datasets. The methods compared include the two-stage detector Faster R-CNN, the single-stage detectors YOLOv8 and YOLOv11, with YOLOv11 serving as the baseline network. SSRetinex+YOLOv11, StarNet, YOLA, ADHI-YOLO, and MANet are selected as representative comparison methods. SSRetinex+YOLOv11 first uses self-supervised Retinex decomposition to improve low-light images before performing object detection; StarNet establishes high-dimensional feature representation through element-wise multiplication; YOLA focuses on lightweight illumination-invariant feature learning; ADHI-YOLO introduces attention SPPF and a dynamic detection head; and MANet employs a multi-scale cascade extraction and adaptive feature fusion mechanism. Considering the difficulty in obtaining paired supervised samples in low-light enhancement tasks, the experiment selects SSRetinex, which has self-supervised learning capabilities, as the image enhancement comparison method. Finally, YOLOv11 with the introduction of a priori-guided dual-domain shallow feature reconstruction module will be uniformly tested to evaluate the overall performance of the method of the present invention in complex low-light target detection tasks.

[0091] Table 3. Results of the ExDark comparative experiment

[0092] Table 3 presents the quantitative results of each detection method on the ExDark dataset. Faster R-CNN's mAP50 and mAP50:95 are 50.1% and 25.0%, respectively, lower than other comparative models. The YOLO series demonstrates better detection capabilities on this dataset, with the YOLOv11 baseline model achieving an mAP50 of 74.1% and an mAP50:95 of 50.7%. While StarNet, YOLA, ADHI-YOLO, and MANet have made improvements in feature modeling or network structure, their detection accuracy still does not surpass the YOLOv11 baseline. Using SSRetinex to pre-enhance low-light images improves the detection results of YOLOv11 to some extent, but the image enhancement process may simultaneously amplify noise or alter local textures, thus interfering with subsequent feature extraction. The method in this invention does not rely on independent image enhancement preprocessing but directly reconstructs low-light shallow features at the network entry point, ultimately achieving an mAP50 of 75.1% and an mAP50:95 of 51.1%, both the highest values ​​in the table. The above results demonstrate that the method of the present invention can improve the detection accuracy and stability in complex low-light scenes while maintaining the target detection capability.

[0093] Table 4. Results of the DarkFace comparative experiment

[0094] The test results on the DarkFace dataset are consistent with the conclusions of the ExDark dataset. Table 4 shows that among the two baseline models, YOLOv8 and YOLOv11, YOLOv11 exhibits superior detection performance, achieving an mAP50 of 65.2%. The mAP50 of StarNet, YOLA, ADHI-YOLO, and MANet mainly falls between 64% and 66%, but they do not show a significant advantage in overall accuracy and detection stability. Combining SSRetinex with YOLOv11 also did not further improve detection metrics, indicating that in nighttime face scenes, explicit image enhancement may simultaneously amplify noise and produce artifacts, thus interfering with target feature extraction. The method of this invention achieves an mAP50 of 66.6% and an mAP50:95 of 33.9% on the DarkFace dataset, representing improvements of 1.4 and 0.8 percentage points respectively compared to the YOLOv11 baseline, with a recall of 53.1%. These results demonstrate that the proposed method achieves a good balance between detection accuracy, target detection rate, and adaptability to complex lighting conditions. Combined with experimental results on the ExDark dataset, it can be seen that the proposed shared prior-driven dual-domain first-layer reconstruction mechanism delivers stable gains in various low-light scene types.

[0095] To further observe the actual detection performance of each method under complex low-light conditions, Figure 8 Visualization results of YOLOv11, StarNet, YOLA, ADHI-YOLO, MANet, and the method of this invention on the ExDark dataset are presented, where Ground-Truth represents the ground truth annotations, and different colored detection boxes correspond to different target categories.

[0096] from Figure 8 It can be seen that YOLA has a weak response to small-scale targets under insufficient illumination or local strong light interference, and some detection boxes also have positional deviations. Although ADHI-YOLO can enhance the response to targets in dark areas, it may still experience detection fluctuations and false detections in extremely low illumination scenes. StarNet is more sensitive to bright areas and reflective edges, easily misidentifying them as targets, and its detection stability for small targets in dark areas is insufficient. MANet also exhibits target omissions and background false responses in extremely dark scenes. YOLOv11 performs relatively stably overall, but when the foreground and background brightness are similar or the target contrast is low, a small number of missed detections and false detections may still occur. In contrast, the method of this invention can enhance the target edge and texture response by leveraging local variance priors and improve the expression of dark area features through color decoupling and compensation, thus performing more stably in target localization, weak target detection, and highlight interference suppression. Its detection boxes have a higher degree of overlap with the real target area and also have better recognition capabilities for targets in shadow areas and low-contrast areas. Figure 8The visual effects reflected corroborate the quantitative results in Table 3.

[0097] Based on the quantitative and qualitative results from the ExDark and DarkFace datasets, the method of this invention can effectively improve detection performance under conditions of weak texture, small-scale targets, localized brightness, and non-uniform illumination, and outperforms comparative methods in terms of detection accuracy, localization stability, and false detection suppression. This method can enhance the visual perception reliability of quadrupedal inspection robots in underground spaces, nighttime inspections, and complex unstructured environments, providing a stable feature base for real-time target detection under low-light conditions.

[0098] Experiment 3: Generalization analysis.

[0099] Although haze and low light levels represent different types of visual degradation, both weaken the difference between the target and the background, leading to reduced edge response, diminished texture information, and indistinct features of small targets. Therefore, this invention uses the RTTS dataset to test the model's adaptability across degraded scenarios. This dataset includes various real-world foggy environments such as urban roads, highways, traffic intersections, and rural roads. Images in this dataset commonly exhibit reduced visibility, unclear target outlines, complex background interference, and local occlusion, placing high demands on the detection of distant, small-scale, and densely packed targets. Therefore, RTTS can be used to verify the transferability and generalization performance of low-light target detection methods under non-low-light degradation conditions.

[0100] Table 5 Results of RTTS Comparison Experiment

[0101] Table 5 presents the test results of each method on the RTTS foggy dataset. The method of this invention achieves an mAP50 of 83.3%, higher than comparable models such as YOLOv8, YOLA, and StarNet; its mAP50:95 is 62.2%, on par with the best result in the table, indicating that this method has relatively stable localization capabilities under different cross-union thresholds. Its recall reaches 76.2%, standing out among various comparative methods, demonstrating that the model can detect more targets and effectively reduce missed detections under foggy and reduced visibility conditions. Unlike methods that rely on independent image enhancement processing, this invention directly reconstructs low-quality visual features at the network entry point, enabling the model to learn feature representations that possess both structural stability and degradation adaptability. Although the model is primarily trained on low-light scenes, it still maintains high detection accuracy and recall on the RTTS dataset, demonstrating good cross-degradation transfer capabilities.

[0102] To further observe the actual detection performance of the model in low-visibility environments, Figure 9This paper presents qualitative comparison results between the method of this invention and YOLA, ADHI-YOLO, StarNet, MANet, and YOLOv11 on the RTTS dataset. In multi-object road scenes, some comparison methods are easily affected by fog, complex backgrounds, and changes in target scale, resulting in phenomena such as detection box offset, false background detection, and omission of small targets at a distance. The method of this invention can more completely cover targets of different scales, and the predicted boxes have higher consistency with the ground truth annotations. For low-contrast targets and distant targets with weak features, comparison models may identify fog textures or background responses as targets, while the method of this invention can more accurately preserve the target contours and complete boundary regression. In scenes with both dense fog and low illumination, the method of this invention can still output relatively compact and reasonable detection boxes.

[0103] Based on experimental results from the ExDark and RTTS datasets, the local variance prior-guided illumination-invariant feature generation module constructed in this invention can enhance edge and texture responses based on local brightness fluctuations and reduce the impact of non-uniform illumination on shallow features. The color decoupling and compensation feature generation module compensates for color features by separating chromaticity and intensity information and introducing brightness priors, enabling a more complete expression of the color, structure, and details of targets in dark areas. After reconstruction by the dual-domain feature reconstruction unit, both types of features provide a more stable input representation for subsequent detection networks. Compared with the YOLOv11 baseline, the method of this invention improves mAP50 by 1.0 and 0.9 percentage points on the ExDark and RTTS datasets, respectively. These results demonstrate that the proposed method not only improves target detection performance in complex low-light environments but also adapts to other degradation conditions such as fog, low visibility, and weak features, providing reliable visual detection support for real-time environmental perception of quadrupedal intelligent inspection robots in underground spaces, nighttime inspections, and complex unstructured scenes.

Claims

1. A low-light target detection method for a brightness-guided dual-domain reconstruction quadrupedal orthotic intelligent robot, characterized in that, Specifically, the steps include the following: Step 1: Construct a priori-guided dual-domain shallow feature reconstruction module and replace the first Conv module at the input of the YOLOv11 backbone network with this module, while retaining the rest of the original YOLOv11 backbone, neck and detection head structures to build a low-light target detection model; The prior-guided dual-domain shallow feature reconstruction module includes a shallow feature generation unit and a dual-domain feature reconstruction unit; the shallow feature generation unit includes a parallel local variance prior-guided illumination-invariant feature generation module and a color decoupling and compensation feature generation module. Step 2: Collect and organize real low-light scene images, complete annotation and size standardization, and establish a low-light target detection dataset; Step 3: Input the images from the low-light target detection dataset into the low-light target detection model, and design a combined loss function to complete the model optimization training; Step 3 specifically includes the following sub-steps: Step 3.1: Input the images from the low-light target detection dataset into the shallow feature generation unit. The illumination-invariant feature generation module and the color decoupling and compensation feature generation module, guided by local variance prior, process the input images simultaneously. The illumination-invariant feature generation module, guided by local variance prior, outputs illumination-invariant features and a brightness map. The brightness map serves as the brightness prior for the color decoupling and compensation feature generation module. The color decoupling and compensation feature generation module combines the brightness prior to output color compensation features enhanced by the brightness prior. The illumination-invariant feature generation module guided by local variance prior includes a parallel candidate feature generation network and a local variance prior guiding network; it also includes a gated residual fusion module. The candidate feature generation network is used to extract attention and enhance candidate features. The initial illumination invariant characteristics are obtained through inter-channel division. ; The local variance prior guidance network is used to calculate the local variance map prior from the input image. ; The gated residual fusion module will use the local variance map as a priori. As weights for attention enhancement candidate features Invariant features with initial illumination Adaptive fusion is performed to obtain illumination-invariant features. , means as follows: ； The local variance prior guidance network processes the input image as follows: First, the input image The R, G, and B channels are taken at their maximum values ​​at each pixel location to obtain a single-channel luminance map. , means as follows: In the formula, and These represent the horizontal and vertical coordinates of the image pixel positions, respectively. Indicates in the channel c Take the maximum value from the middle. , These represent the height and width of the input image, respectively. Indicates the input image in channels c Pixel value at; Meanwhile, brightness map As a priori output of brightness; Then, the brightness map Intra-channel feature mapping is performed using a 5×5 depthwise convolution kernel to obtain the brightness feature map. , is represented as: In the formula, Indicates a depthwise convolution operation; Next, a depthwise convolution with a kernel size of 5×5 is used to aggregate neighborhood information within each channel, and the ith value is calculated. Each channel is located at coordinate position The local mean at point is expressed as: In the formula, Indicates the first Each channel is located at coordinate position Local mean at that location Indicates The 5×5 neighborhood centered on the center Represents brightness feature mapping In the Coordinate position in each channel The characteristics of the place, u and v This represents the horizontal and vertical offsets within the local convolution window; Based on this, brightness feature mapping In the Coordinate position in each channel Features of the place With the Each channel is located at coordinate position Local mean at The squared deviations are then aggregated using a 5×5 depthwise convolution to obtain the coordinate positions. Prior local variance plot , means as follows: ； The candidate feature generation network processes the input image as follows: First, the input image The results are processed by 3×3 convolutional layers and 5×5 convolutional layers respectively to obtain local responses at different scales. The two results are then merged along the channel direction to obtain multi-scale candidate features. Secondly, multi-scale candidate features are input into the channel-space joint attention mechanism: first, global average pooling, global max pooling, and shared fully connected mapping are used to obtain channel weights; then, average pooling, max pooling, and 7×7 convolutions are performed on the channel-recalibrated features to obtain spatial weights; finally, attention-enhanced candidate features are output. ; Subsequently, illumination-invariant modeling was performed on the attention-enhancing candidate features. The impact of global illumination scale variations was mitigated through numerical stabilization and adjacent channel ratio division. Specifically, this included: To avoid zero denominators in channel ratio division, attention-enhanced candidate features are used. The numerical stabilization process is performed beforehand, as shown below: In the formula, It is a constant. For symbolic functions, To enhance attention to candidate features The processed value; Attention-enhanced candidate features after function processing; By performing proportional division on adjacent channels in a cyclic manner and then stitching them together along the channel dimension, the initial illumination-invariant feature is obtained, as shown below: In the formula, Indicates the first i Attention-enhanced candidate features for each channel after processing by a function. This indicates a splicing operation at the channel level. This indicates that the initial illumination remains unchanged. The color decoupling and compensation feature generation module includes an HVI color decoupling unit, a color compensation feature extraction unit, and a brightness prior-guided enhancement unit connected in sequence. The image processing procedure of the color decoupling and compensation feature generation module on the input low-light target detection dataset is as follows: First, the input image is a standard three-channel RGB image. Input to HVI color decoupling unit: Calculate the input image brightness value saturation Hue Let the maximum and minimum channel responses at the same pixel be respectively... T and m Then saturation Calculate using the following formula: In the formula, It is a constant; Indicates in the channel c Take the minimum value from the middle. Introducing a color sensitivity adjustment factor joint saturation Hue Polarization mapping is performed to obtain hue components. and vertical components Define intensity components I And assign the brightness value to the intensity component. I This constitutes the HVI decoupling representation, as shown below: In the formula, Represents the hue value of a single pixel; Among them, the color sensitivity adjustment factor Calculate using the following formula: In the formula, This indicates the color sensitivity adjustment parameter; Secondly, the color compensation feature extraction unit encodes the hue component and the vertical component: , Color compensation features are obtained after splicing and chromaticity residual coding. At the same time, the intensity component I Strength constraint features are obtained through strength constraint encoding. , means as follows: In the formula, , They represent and Convolution operation; where the chroma residual encoding consists of 3×3 convolutions and the intensity constraint encoding consists of 5×5 convolutions; Subsequently, the luminance prior-guided enhancement unit analyzed the luminance map. Color compensation features and strength constraint features Joint fusion is performed: After spatial size matching of the luminance map with the color compensation features and intensity constraint features, they are stitched together along the channel dimension to obtain the color compensation features with luminance prior enhancement. , means as follows: In the formula, This indicates a splicing operation along the channel dimension; Step 3.2: Input the illumination-invariant features and the brightness-prior-enhanced color compensation features into the dual-domain feature reconstruction unit to generate the first-layer reconstructed features with the same output size and number of channels as the original first Conv module of YOLOv11; Step 3.3: Input the reconstructed features from the first layer into the subsequent YOLOv11 network, and complete the model optimization training by combining loss functions to obtain the trained low-light target detection model; Step 4: Input the low-light image to be detected into the optimized and trained low-light target detection model, and output the target category, confidence score and location results.

2. The low-light target detection method for a brightness-guided dual-domain reconstruction quadrupedal anthropomorphic intelligent robot according to claim 1, characterized in that, Step 2 specifically includes: Real-world low-light scene images are collected, and the raw image data is cleaned and filtered to remove blurry, overexposed, and invalid samples without targets. The image resolution is standardized to 1280×720 pixels, and the data is regularized and standardized to obtain a low-light target detection dataset for model training and testing.

3. The low-light target detection method for a brightness-guided dual-domain reconstruction quadrupedal anthropomorphic intelligent robot according to claim 1, characterized in that, The dual-domain feature reconstruction unit includes channel splicing units connected in series. Convolutional layers, batch normalization layers, and the SiLU activation function; The output illumination-invariant feature Color compensation features with prior brightness enhancement Input a dual-domain feature reconstruction unit and output a first-layer reconstructed feature that combines illumination robustness and color perception capability. , means as follows: In the formula, This indicates a channel splicing operation. This indicates a batch normalization operation. This represents the SiLU activation function. express Convolution operation.

4. The low-light target detection method for a brightness-guided dual-domain reconstruction quadrupedal anthropomorphic intelligent robot according to claim 1, characterized in that, The combined loss function consists of bounding box regression loss. Targeted loss and classification loss The weighted composition is represented as follows: In the formula, , , The bounding box regression loss is respectively Targeted loss and classification loss The weight, Denotes the combined loss function; where, , , All numbers are positive and satisfy the following conditions: ； The bounding box regression loss , means as follows: In the formula, This represents an improved IoU loss calculated based on the overlap between the predicted bounding box and the ground truth bounding box. This represents the distributed regression loss; The internal adjustment coefficient of distributed regression loss in bounding box regression loss; The targeted loss , means as follows: In the formula, The target score predicted by the model. For soft labels related to IoU quality; Indicates loss of quality focus; The classification loss is similar to the output shape based on the category of the detection head. Use either multi-class cross-entropy or class-by-class binary cross-entropy summation: Using a multi-class cross-entropy approach: Samples are labeled with hot tags, with the target class label set to 1 and other class labels set to 0, as shown below: In the formula, Indicates the total number of categories. Indicate category The true label, To output the category to which the sample belongs The predicted probability; The classification loss is obtained by summing the binary cross-entropy values ​​for each category: the binary cross-entropy is calculated separately for each category and then summed to obtain the total classification loss, as shown below: 。

Images