UMFNet-YOLO-based joint detection algorithm under low light condition

Abstract

The invention relates to the field of target detection, in particular to a UMFNet-YOLO-based joint detection algorithm under a low-light condition, which adopts a multi-scale decomposition and high-frequency-low-frequency feature collaborative enhancement strategy, separates image details from global features through high / low-pass filtering, realizes multi-resolution feature complementation in combination with dynamic weight distribution, and improves the detection accuracy. Target contour and texture information degraded in severe weather can be effectively recovered; a parallel channel compression and spatial semantic enhancement module is designed, rain and fog noise interference is suppressed through cross-dimensional adaptive feature screening, and significance expression of a key target is enhanced; an image enhancement network UMFNet is cascaded with YOLOv8, and strong correlation between feature expression and a detection task is ensured by synchronously optimizing an enhancement module and detecting network parameters through back propagation. Compared with a baseline YOLOv8 detector, the method has the advantages that the average precision (mAP50) is obviously improved, and the method is obviously superior to a traditional enhancement-detection separation scheme. The method can be applied to robust target detection scenes in complex environments such as automatic driving and security monitoring.

Description

Technical Field

[0001] The present invention relates to the fields of computer vision and target detection, and in particular to a joint detection algorithm under low-light conditions based on UMFNet-YOLO. Background Art

[0002] The key challenge of low-light object detection is the loss of effective features caused by image degradation (noise, blur, low illumination). Existing solutions have the following deficiencies: (1) Limitations of traditional image enhancement methods: Histogram equalization and Retinex theory: They enhance brightness by global contrast stretching, but cannot distinguish between noise and effective signals, resulting in overexposure of details or amplification of noise; (2) Manual filters (such as bilateral filtering): They rely on manual parameter adjustment and have poor adaptability to dynamic low-light scenes; (3) Insufficient existing deep learning models: Global enhancement models: They focus on the visual effect of the entire image and ignore the feature enhancement of local key areas (such as object edges), resulting in a high misjudgment rate of the detection network; (4) Single-scale feature extraction: They have limited ability to capture details of multi-scale targets (small objects at a distance, occluded targets); (5) Existing joint training methods: They use fixed parameter enhancement modules, which cannot dynamically adapt to different light intensities and are prone to introducing feature offsets (experiments show that their false detection rate increases by 15% in low-light mixed scenes); (6) Information attenuation of cascade architectures: In existing cascade models (such as DE-YOLO), the detection network and enhancement modules are trained independently, resulting in feature transfer distortion (for example, the loss rate of high-frequency details in cross-module transfer is more than 20%). Therefore, existing technologies fail to effectively address the issues of local feature degradation and cross-module compatibility, leading to bottlenecks in detection accuracy. There is an urgent need to research a new architecture that can dynamically fuse multi-scale features, suppress noise, and coordinate optimization with the detection network. Summary of the Invention

[0003] To address the issues of feature blurring, noise interference, and decreased detection accuracy caused by image quality degradation in harsh environments such as rain, fog, and low light, this paper provides a joint detection algorithm for low-light conditions based on UMFNet-YOLO. This algorithm is a UMFNet-YOLO joint detection framework that integrates multi-scale feature enhancement and dynamic attention optimization. It mainly includes: S1: Obtain an image dataset and preprocess it. After preprocessing, it is divided into a training set and a test set according to a set ratio; S2: Based on the UNet encoder structure, multi-scale features are extracted from the images in the training set and the test set step by step. The high-resolution features and adjacent low-resolution features in the multi-scale features are input into the frequency adaptive fusion module to separate the high-frequency and low-frequency features. S3: Adopting multi-scale decomposition and high-frequency-low-frequency feature collaborative enhancement strategy, image details and global features are separated by high / low-pass filtering, and high-frequency-low-frequency complementary enhancement processing is performed on the high-frequency and low-frequency features separated in step S2 to obtain cross-scale fusion features; S4: Design a parallel channel compression and spatial semantic enhancement module, filter out fusion features that suppress rain and fog noise interference through cross-scale fusion features, and perform SCCA optimization on the fusion features; S5: Perform transposed convolution upsampling on the high-level features optimized by SCCA, and concatenate them with the low-level features step by step across jumps. Use lightweight convolution blocks to fuse multi-scale details and output an enhanced image. S6: UMFNet-YOLO is constructed by combining UMFNet and YOLOv8. The self-attention mechanism is integrated into the Neck part of YOLOv8 to enhance the association of multi-scale objects. S7: The enhanced image is used to train the UMFNet-YOLO. During the training process, the UMFNet and YOLOv8 parameters in the UMFNet-YOLO are synchronously adjusted to finally obtain a detection model adaptive to low-light conditions. The target position, target category, and confidence level output by the detection model are subjected to NMS processing to generate a robust detection result.

[0004] A computer device includes a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the above method.

[0005] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the above method.

[0006] A computer program product includes a computer program or instructions, which implement the steps of the above method when the program or instructions are executed by a processor.

[0007] The technical solution provided by this invention offers the following beneficial effects: It utilizes a multi-scale decomposition and high- and low-frequency feature collaborative enhancement strategy, separating image details from global features through high- and low-pass filtering. Combined with dynamic weight allocation, it achieves multi-resolution feature complementarity, effectively restoring degraded object contours and texture information in inclement weather. It also designs a parallel channel compression and spatial semantic enhancement module, suppressing rain and fog noise interference through cross-dimensional adaptive feature filtering, thereby enhancing the saliency of key objects. Furthermore, it employs an end-to-end joint training architecture: cascading the image enhancement network UMFNet with the YOLOv8 detector, and simultaneously optimizing the enhancement module and detection network parameters through backpropagation to ensure strong correlation between feature representation and detection tasks. This method achieves significant improvements in mean average precision (mAP50) compared to the baseline YOLOv8 detector in the EXDark low-light dataset and synthetic rain and fog scenes, significantly outperforming traditional enhancement-detection separation schemes. This algorithm is suitable for robust object detection in complex environments such as autonomous driving and security surveillance. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] The present invention will be further described below with reference to the accompanying drawings and embodiments, in which: Figure 1 This is a flowchart of a joint detection algorithm based on UMFNet-YOLO under low-light conditions in an embodiment of the present invention; Figure 2 Schematic diagram of the overall architecture of the multi-scale feature fusion network (UMFNet-YOLO) in an embodiment of the present invention; Figure 3 Schematic diagram of the structure of the encoder-decoder backbone network in an embodiment of the present invention; Figure 4 is a schematic diagram of a FAF according to an embodiment of the present invention; Figure 5 Schematic diagram of SCCA in an embodiment of the present invention. DETAILED DESCRIPTION

[0009] In order to have a clearer understanding of the technical features, purposes and effects of the present invention, specific embodiments of the present invention are now described in detail with reference to the accompanying drawings.

[0010] Example 1 Please refer to Figure 1 , Figure 1 This is a flowchart of a joint detection algorithm based on UMFNet-YOLO under low-light conditions in an embodiment of the present invention, specifically including: S1. Image Input and Noise Preprocessing: Preprocess the raw noisy input image under low-light conditions (including low-light and rain-fog interference scenes). This includes the following operations: normalize the input image to adjust the pixel mean and variance to a preset range; scale the resolution to a predetermined size (e.g., 640×640 pixels) and perform preliminary noise suppression using bilateral filtering to reduce subsequent feature interference.

[0011] S2. Multi-scale Feature Encoding and Frequency Decomposition: Based on the UNet encoder structure, multi-scale features are extracted from the preprocessed image step by step. Four layers of feature maps {p1, p2, p3, p4} with different resolutions are generated through cascaded convolutional layers and maximum pooling operations. The resolution is reduced to 1 / 2, 1 / 4, 1 / 8, and 1 / 16, and the number of channels is increased to 8, 16, 32, and 64. The high-resolution features (such as p1) and the adjacent low-resolution features (such as p2) are input into the frequency adaptive fusion module (FAF) to separate the high-frequency and low-frequency features.

[0012] Frequency adaptive feature fusion low light adaptability: Aiming at the high noise and low contrast characteristics of low light images, a frequency adaptive fusion module (FAF) is proposed to realize the sorting and fusion of high-frequency and low-frequency features.

[0013] S3. A multi-scale decomposition and high- and low-frequency feature collaborative enhancement strategy is employed. High- and low-pass filtering is used to separate image details from global features. Dynamic weight allocation is then used to achieve multi-resolution feature complementation, effectively restoring target contours and texture information degraded in inclement weather. The multi-scale features from step S2 are then subjected to high- and low-frequency complementary enhancement to generate cross-scale fused features: a. High-frequency feature extraction: High-pass filtering is used to convolve the high-resolution feature map to separate edges and texture details. b. Low-frequency feature enhancement: Low-resolution feature maps are downsampled and filtered using a low-pass filter to extract global structural information. c. Dynamic weight allocation: High- and low-frequency weight matrices are adaptively generated based on feature content. d. Cross-scale fused feature acquisition: The weighted high- and low-frequency features are concatenated and refined through multi-scale convolutional layers (3×3, 5×5, and 7×7).

[0014] S4. Design a parallel channel compression and spatial semantic enhancement module. Through cross-scale feature fusion screening, suppress the interference of rain and fog noise, enhance the saliency of key targets, and optimize the spatial-channel joint attention mechanism (SCCA) of the fused features.

[0015] Multi-scale perception enhancement: Construct a spatial-channel joint attention mechanism (SCCA) to optimize the coordinated expression of local details and global semantics.

[0016] The parallel channel compression and spatial semantic enhancement module suppresses rain and fog noise interference through cross-scale adaptive feature filtering, enhancing the saliency of key targets. Cross-dimensional optimization is performed on the fused features: a. Channel attention branch: Global average pooling (GAP) is performed on the feature map to generate a channel description vector. This is then processed through a one-dimensional convolution and activation function to generate a channel weight matrix. This weight matrix is multiplied channel by channel with the original features to enhance the representation of key feature channels. b. Spatial attention branch: Grouped average pooling is performed along the height and width of the feature map, and spatial convolution is used to generate a spatial weight matrix to suppress background noise.

[0017] c. Mixed attention output: Affine fuse the channels with the spatial weight matrix to obtain the optimized feature map.

[0018] S5. Feature decoding and enhanced image generation: Image resolution is gradually restored through the UNet decoder: The high-level features optimized by SCCA are upsampled by transposed convolution and then spliced with the low-level features step by step across jumps. Lightweight convolution blocks are used to fuse multi-scale details and output an enhanced image (with the same resolution as the input).

[0019] S6. Joint object detection based on YOLOv8: The generated enhanced image is fed into a multi-scale feature fusion network (UMFNet-YOLO). The UMFNet-YOLO is a cascade of UMFNet and YOLOv8. Integrate the self-attention mechanism in the neck part of YOLOv8 to strengthen the multi-scale object association; A decoupled head structure is used to separate the classification and regression branches to predict the target category and location box (i.e., the target location).

[0020] S7. End-to-end cascade architecture, joint training optimization, and synchronous adjustment of UMFNet and YOLOv8 parameters. This joint training strategy deeply integrates UMFNet with the YOLOv8 detection network, improving the robustness of UMFNet-YOLO under complex lighting conditions. a. Composite loss function design: This includes the perceptual loss (LPIPS) of the enhancement network and the CIoU classification and regression loss of the detection network. b. Backpropagation strategy: During training, the detection error gradient is backpropagated to the UMFNet parameters to ensure alignment of the enhanced features with the detection task. c. Two-stage optimization mechanism: Pre-training stage: freeze YOLOv8 weights and train UMFNet alone until convergence; Joint fine-tuning stage: Unfreeze the detection network parameters and optimize the overall model until the loss is stable, finally obtaining a detection model that is adaptive to low-light conditions.

[0021] The target position, target category, and confidence level output by UMFNet-YOLO are processed with non-maximum suppression (NMS) to generate robust detection results for application in autonomous driving or security monitoring systems.

[0022] The structure diagram of the multi-scale feature fusion network (UMFNet-YOLO) is as follows Figure 2 As shown in the figure, it is composed of a cascade deployment of UMFNet and YOLOv8 detection models for low-light target detection. The specific structure includes the following parts: A. Encoder-Decoder Backbone Network ( Figure 3 ) (1) Encoding stage (feature extraction): Input low-light image , after 4 levels of downsampling: Downsampling block 1: 3×3 convolution (stride 2, padding 1) + batch normalization (BN) + LeakyReLU, output feature map

[0023] Downsampling blocks 2 to 4: gradually increase the number of channels (16 → 32 → 64), and reduce the resolution proportionally to the original Figure 1 / 16. Each level contains residual connections to avoid gradient disappearance.

[0024] (2) Decoding stage (feature reconstruction): The resolution is restored by bilinear upsampling. Each decoder receives the same-scale jump connection features from the encoder and passes through the FAF module (multi-scale fusion) and SCCA module (feature optimization) in sequence: Upsampling block 1: Deconvolution (kernel 3×3, stride 2) expands the feature from 64 channels to 32 channels and increases the resolution to 1 / 8.

[0025] FAF module: fuses the high-frequency (details) and low-frequency (structure) features of the current level (see below for details).

[0026] SCCA module: Joint channel and spatial attention enhancement key areas (details are given below).

[0027] 2. Frequency Adaptive Fusion Module (FAF) Design ( Figure 4 ) The FAF module enhances effective features and suppresses noise interference through frequency division processing: Step 1: Frequency Decomposition 1.1 High-frequency feature extraction: The input features are filtered by high-pass filtering (HPF) to capture edge details. The formula is:

[0028] in It is a Laplace sharpening kernel that enhances edge response.

[0029] 1.2 Low-frequency feature extraction: input features The global structure is extracted through low-pass filtering (LPF). The process is as follows: Average pooling (kernel 3×3, stride 2) downsamples to 1 / 4 size. Depthwise separable convolution (kernel 5×5, depth magnification 4) smooths noise. Bilinear upsampling restores to original size. Step 2: Adaptive feature fusion 2.1 Multi-scale convolution branch: Parallel convolution kernels (1×1, 3×3, 5×5) are used to extract multi-scale features. The formula is:

[0030] is the convolution parameter corresponding to the kernel size, and the output feature , .

[0031] 2.2 Dynamic Weight Generation: Generating Multi-branch Weights through Channel Attention: Splicing features of each branch

[0032] The compressed excitation network (SE) calculates branch importance:

[0033] GAP is global average pooling, and is the fully connected layer, is ReLU activation, and σ is Sigmoid normalization.

[0034] 2.3 Weighted fusion output: Aggregate each branch by weight:

[0035] 3. Spatial Channel Joint Attention (SCCA) Mechanism ( Figure 5 ) The SCCA module optimizes feature representation through a dual attention path: A. Channel Attention Path Global Information Compression: Input Features After global average pooling .

[0036] Dynamic weight generation: Generate channel weights through a multi-layer perceptron (MLP):

[0037] in, , ,r is the compression ratio (default r=16).

[0038] B. Spatial Attention Path Multi-directional feature extraction: Average pooling along the width axis , width weights are generated by 1×1 convolution + Sigmoid.

[0039] Average pooling along the height axis , 1×1 convolution + Sigmoid generates height weights.

[0040] Spatial weight fusion:

[0041] For the broadcast addition operation, the final spatial weight matrix . C. Joint Attention Generation Affine fusion: nonlinear fusion of channels and spatial weights:

[0042] in are learnable parameters.

[0043] Feature Enhancement: Output Features .

[0044] 4. Joint Training Strategy A. Loss Function Composition Detection loss : The loss defined by YOLOv8, including bounding box CIoU loss, classification cross entropy loss, and confidence loss:

[0045] in =0.05, =0.5, =1.0.

[0046] Enhancement Loss :

[0047] To enhance the image, is the normal lighting reference image, .

[0048] Noise suppression regularization term :

[0049] is the image gradient operator, Suppresses high-frequency noise. B. Training Process Data preprocessing: Input images are normalized to [0,1], Mosaic enhancement (first 240 epochs), random flipping (probability 0.5), and color perturbation (brightness ±20%, contrast ±15%) are applied. Staged training: Stage 1 (50 epochs): Freeze the YOLOv8 backbone and train only the UMFNet parameters, learning rate .

[0050] Phase 2 (200 epochs): Jointly fine-tune the entire system, and anneal the learning rate cosine to .

[0051] 5. Technological advantages Verified by the EXdark dataset, the present invention has the following significant advantages over the existing technology: Improved accuracy: mAP@50 increased by 2.04% to 70.1%, and the false positive rate decreased by 12.6% (compared to the original YOLOv8).

[0052] The detection success rate in extreme low light (brightness <5 lux) is 89.7%, which is better than EnlightenGAN (75.3%) and MAET (83.2%).

[0053] Efficiency optimization: Single-frame inference speed of 45 FPS (NVIDIA A800) meets real-time detection requirements.

[0054] The number of model parameters only increases by 4.7% (compared to cascaded DE-YOLO).

[0055] Enhanced robustness: Under complex degradation conditions such as rain, fog, and motion blur, mAP@50 remains ≥65.3% (the comparison model drops to ≤58%).

[0056] 1. System construction and hardware configuration Development environment: Operating system: Ubuntu 20.04 LTS.

[0057] Deep learning framework: PyTorch 1.12 + CUDA 11.6.

[0058] GPU configuration: NVIDIA A800 80GB × 4 (distributed data parallel training).

[0059] Deployment hardware: Edge device: Jetson AGX Xavier (highest power mode).

[0060] Memory requirement: A minimum of 4GB of video memory is required to run real-time inference.

[0061] 2. Detailed steps for network construction Step 1: UMFNet image enhancement network construction Encoder construction: The following hierarchy is used: Level-1: Input layer → 3×3 convolution (channel expansion: 3→8) → LeakyReLU (α=0.1); Level-2: Residual module × 3 (the number of channels remains 16, and each residual block contains two parallel convolution kernels: 3×3 and 5×5); Level-3: Atrous convolution layer (dilation rate = 2) → batch normalization → Swish activation; Level-4: Depthwise separable convolution → max pooling (window 2×2) → channel attention embedding.

[0062] Decoder construction: Each layer cascade performs the following operations: bilinear upsampling (scaling factor = 2) → FAF module → SCCA module → cross-layer feature concatenation.

[0063] The channel number change path is: 64 → 32 → 16 → 8, and the final output dimension is aligned with the input image.

[0064] Step 2: Check network integration Connect the YOLOv8 detection head at the output of UMFNet and make the following adjustments: The input channels of the detection head are modified to match the UMFNet output (the original 640 channels are adjusted to 256 channels).

[0065] Anchor size re-clustering generation: A new size set {[23,27], [46,39], [77,66]} is obtained based on the EXdark dataset.

[0066] SCCA module key parameters: Channel attention compression ratio: r=16 (full connection layer dimension change C→C / r→C).

[0067] Spatial attention direction weight generation: Height branch: average pooling along the Width dimension → 1×1 convolution (channel reduction ratio 1 / 4); Width branch: average pooling along the Height dimension → 1×1 convolution (channel preservation).

[0068] The fusion coefficient β is initialized to 0.5 and automatically adjusted during training.

[0069] 4. Training process parameter configuration Data preprocessing process: Degradation simulation: A degradation model is applied to the normal illumination image, where 𝒫 is a random patch noise generator to simulate regional information loss in low light.

[0070] Online enhancement: Mosaic enhancement ratio: applied in the first 200 epochs, mixing 4 images randomly; Color perturbation range: HSV space ±20% perturbation; Random Erase: 10% chance of erasing a 30×30 pixel area.

[0071] Optimizer parameter group:

[0072] 5. Deployment optimization strategy Model lightweight processing: Channel pruning: After performing a sensitivity analysis on the validation set, the number of Level-4 channels was pruned down from 64 to 48 (accuracy loss <0.5%). The YOLOv8 detection head channel is reduced by 30%.

[0073] Quantization deployment: Post-training quantization (PTQ): FP32→INT8, using the EMA correction algorithm; Dynamic range calibration: 500 images from the exdark_val dataset are used for calibration.

[0074] Inference acceleration optimization: TensorRT engine configuration: <cpp> config->setFlag(BuilderFlag::kFP16); config->setMemoryPoolLimit(MemoryPoolType::kWORKSPACE, 1<<30); config->setProfilingVerbosity(ProfilingVerbosity::kDETAILED); Multi-threaded processing: input queue-processing engine-output queue three-stage pipeline parallel 6. Exception handling mechanism Light failure detection: Real-time calculation of image mean μ and standard deviation σ; If μ<30 and σ<10, it is judged as an extreme low-light scene, triggering dual-model integrated reasoning.

[0075] Performance regression protection: Store the historical 10-frame detection results on the device; When the confidence level of three consecutive frames drops by more than 15%, the system automatically switches to degradation mode: the input resolution is reduced to 640×480; and the P6 large object detection branch of YOLOv8 is disabled.

[0076] Implementation Verification Verification on 20 night scene categories of the EXDark dataset shows that (the test set has a total of 10,328 images): Accuracy-efficiency balance: 74.9 mAP@50 at an input resolution of 1280×720, and a real-time frame rate of 48 FPS; After quantization, the model runs 210% more efficiently on the Jetson device.

[0077] Module validity verification:

[0078] Robustness in extreme scenarios: Simulating a rainy and foggy environment (adding 10% Gaussian noise): mAP@50 remained at 67.3 (the comparison solution dropped to 58.5); Dynamic blur scene (motion blur kernel 15×15): The detection rate only dropped by 4.2pp (compared to the comparison solution which dropped by 15.7pp).

[0079] Example 2 A computer device includes a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the above method.

[0080] Example 3 A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the above method.

[0081] Example 4 A computer program product includes a computer program or instructions, which implement the steps of the above method when the program or instructions are executed by a processor.

[0082] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.< / cpp>

Claims

1. A joint detection algorithm based on UMFNet-YOLO under low light conditions, characterized by: include: S1: Obtain an image dataset and preprocess it. After preprocessing, it is divided into a training set and a test set according to a set ratio; S2: Based on the UNet encoder structure, multi-scale features are extracted from the images in the training set and the test set step by step. The high-resolution features and adjacent low-resolution features in the multi-scale features are input into the frequency adaptive fusion module to separate the high-frequency and low-frequency features. S3: Adopting multi-scale decomposition and high-frequency-low-frequency feature collaborative enhancement strategy, image details and global features are separated by high / low-pass filtering, and high-frequency-low-frequency complementary enhancement processing is performed on the high-frequency and low-frequency features separated in step S2 to obtain cross-scale fusion features; S4: Design a parallel channel compression and spatial semantic enhancement module, filter out fusion features that suppress rain and fog noise interference through cross-scale fusion features, and perform SCCA optimization on the fusion features; S5: Perform transposed convolution upsampling on the high-level features optimized by SCCA, and concatenate them with the low-level features step by step across jumps. Use lightweight convolution blocks to fuse multi-scale details and output an enhanced image. S6: UMFNet-YOLO is constructed by combining UMFNet and YOLOv8. The self-attention mechanism is integrated into the Neck part of YOLOv8 to enhance the association of multi-scale objects. S7: The enhanced image is used to train the UMFNet-YOLO. During the training process, the UMFNet and YOLOv8 parameters in the UMFNet-YOLO are synchronously adjusted to finally obtain a detection model adaptive to low-light conditions. The target position, target category, and confidence level output by the detection model are subjected to NMS processing to generate a robust detection result.

2. The UMFNet-YOLO-based joint detection algorithm for low-light conditions according to claim 1, wherein: In S1, preprocessing includes: normalizing the image, adjusting the pixel mean and variance to a preset range; uniformly scaling the resolution to a predetermined size, and using bilateral filtering to perform preliminary noise suppression to reduce subsequent feature interference.

3. The UMFNet-YOLO-based joint detection algorithm for low-light conditions according to claim 1, wherein: S3 specifically: S3.1: High-frequency feature extraction: Use a high-pass filter to convolve the high-resolution feature map to separate edge and texture details; S3.2: Low-frequency feature enhancement: The low-resolution feature map is downsampled and filtered through a low-pass filter to extract global structural information; S3.3: Dynamic weight allocation: Adaptively generate high-frequency and low-frequency weight matrices based on feature content; S3.4: Cross-scale fusion: Concatenate the weighted high- and low-frequency features, and refine the fused features through multi-scale convolutional layers.

4. The UMFNet-YOLO-based joint detection algorithm for low-light conditions according to claim 1, wherein: S4 is specifically: S4.1: Channel attention branch: Performs global average pooling on the feature map to generate a channel description vector, which is then processed through a one-dimensional convolution and activation function to generate a channel weight matrix. The weight matrix is multiplied by the original features channel by channel to enhance the expression of key feature channels. S4.2: Spatial attention branch: performs group average pooling along the height and width of the feature map, generates a spatial weight matrix through spatial convolution, and suppresses background noise; S4.3: Mixed attention output: Affine fuse the channels with the spatial weight matrix to obtain the optimized feature map.

5. The joint detection algorithm based on UMFNet-YOLO under low light conditions according to claim 1, characterized in that: S7 specifically: S7.1: Composite loss function design: This includes the perceptual loss of the enhancement network and the CIoU classification regression loss of the detection network. S7.2: Backpropagation strategy: During the training phase, the detection error gradient is backpropagated to the UMFNet parameters to ensure that the enhanced features are aligned with the detection task. S7.3: Two-stage optimization mechanism: Pre-training stage: freeze YOLOv8 weights and train UMFNet alone until convergence; Joint fine-tuning stage: Unfreeze the YOLOv8 detection network parameters, optimize the overall model until the loss is stable, and finally obtain a detection model adaptive to low-light conditions.

6. A computer device comprising a memory, a processor, and a computer program stored in the memory, wherein: The processor executes the computer program to implement the steps of the joint detection algorithm under low-light conditions based on UMFNet-YOLO as described in any one of claims 1-5.

7. A computer-readable storage medium, characterized in that A computer program is stored, and when the program is executed by a processor, the steps of the joint detection algorithm under low-light conditions based on UMFNet-YOLO are implemented.

8. A computer program product, characterized in that The method comprises a computer program or an instruction, which, when executed by a processor, implements the steps of the joint detection algorithm under low-light conditions based on UMFNet-YOLO according to any one of claims 1 to 5.

Images