Low-illumination target detection method

Abstract

The invention belongs to the technical field of image recognition, and provides a low-illumination target detection method, which comprises the following steps of: improving a model structure by taking YOLOv11n as a reference model, and designing a noise frequency band sensing spectrum feedforward neural network module, an aliasing sensing frequency reconstruction up-sampling module and an illumination sensing multi-attention fusion detection head; wherein the noise frequency band perception spectrum feedforward neural network module introduces a frequency domain adaptive modulation and illumination perception double-gating mechanism, the aliasing perception frequency reconstruction up-sampling module introduces a frequency perception and reconstruction mechanism, and the illumination perception multi-attention fusion detection head introduces an illumination perception gating mechanism; using an ExDark low illumination data set to train an improved network model; and performing low-illumination image detection by using the trained model. According to the method, the high reasoning speed can be kept in a low-light environment, and the detection capability and the positioning accuracy of weak visible targets, fuzzy edge targets and small-scale targets are improved.

Description

Technical Field

[0001] This invention relates to the field of image recognition technology, and more specifically to a method for detecting targets in low light conditions. Background Technology

[0002] With the rapid advancement of urbanization and industrialization, the demand for visual perception in low-light environments at night has become increasingly prominent in recent years. Applications such as road security, autonomous driving, and nighttime surveillance are crucial supports for modern social safety and intelligence, and are vital to public safety and industrial development. However, due to complex nighttime lighting conditions, low target visibility, and significant image noise interference, the detection performance of vision systems based on visible light images is generally limited in low-light environments. Issues such as missed detections, false detections, and blurred boundaries directly affect the reliability of monitoring systems and the safety of autonomous driving. Achieving efficient, accurate, and robust target detection under complex lighting conditions has become a current research hotspot and challenge.

[0003] Target detection technology in low-light environments is crucial for improving the perception capabilities of nighttime visual systems. Rapid and accurate target detection in low-light images is a focal point of attention for both academic and industrial communities both domestically and internationally. With the continuous development of deep learning and computer vision technologies, many researchers have proposed various deep learning-based target detection models. Currently, single-stage detectors (such as the YOLO series) have achieved significant results under normal lighting conditions, but research on detection methods for targets in low-light scenarios with complex lighting conditions and weak visibility remains relatively insufficient. Therefore, further exploration of deep learning models based on visual enhancement and feature reconstruction mechanisms in low-light detection tasks is needed.

[0004] Existing low-light detection schemes often employ a two-stage paradigm that separates image enhancement and detection, or only make local improvements to the backbone network, which limits their inference efficiency and feature consistency. In particular, in areas such as upsampling reconstruction, multi-scale feature fusion, and detection head optimization, they generally do not fully consider the frequency domain aliasing, noise band distribution, and the differential impact of lighting conditions on classification and localization tasks in low-light images, leading to the loss of small target features, blurred edges, and performance degradation of the model under extreme low light conditions. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a low-light target detection method to solve the problems of high noise interference, poor feature visibility, and easy missed detection of small targets when detecting low-light images.

[0006] This invention provides a low-light target detection method, comprising: Based on YOLOv11n, the model structure was improved by designing a noise band sensing spectral feedforward neural network module, an aliasing sensing frequency reconstruction upsampling module, and an illumination sensing multi-attention fusion detection head. Among them, the noise band sensing spectral feedforward neural network module introduces a frequency domain adaptive modulation and illumination sensing dual-gating mechanism, the aliasing sensing frequency reconstruction upsampling module introduces a frequency sensing and reconstruction mechanism, and the illumination sensing multi-attention fusion detection head introduces an illumination sensing gating mechanism. The improved network model was trained using the ExDark low-light dataset. Use the trained model to detect low-light images.

[0007] Optionally, the noise band sensing spectral feedforward neural network module performs the following steps in sequence: Channel expansion and spatial context extraction: The input feature map is expanded in channel dimension, and then multi-scale spatial context features are extracted; Frequency domain transformation and band-aware modulation: converting spatial domain features into frequency domain representation, constructing three frequency band mask matrices for low frequency, mid frequency, and high frequency, and performing differentiated modulation on each frequency band component through learnable frequency band weight parameters; Adaptive fusion of lighting: generating lighting perception gating coefficients based on global average pooling and compressed excitation network; Inverse Transform and Noise-Aware Gating: The frequency domain features after modulation and fusion are inversely transformed back to the spatial domain and divided into two equal parts along the channel dimension. The two parts of features are then combined through activation functions and noise gating to achieve noise-sensitive feature selection. Feature integration and output: Compress the features that have undergone noise-aware gating back to the original channel dimension.

[0008] Optionally, the aliasing-sensing frequency reconstruction upsampling module performs the following steps in sequence: Zero-value insertion upsampling operation: Performs zero-value filling upsampling based on regular grid on the input low-resolution feature map. This is achieved by creating an all-zero tensor with a spatial size that is an integer multiple of the original feature map and mapping the input feature values ​​to the corresponding grid positions of the output tensor at fixed intervals, thereby enlarging the spatial size of the feature map to the target multiple. Multi-band depthwise convolutional feature extraction: Perform multi-scale parallel depthwise separable convolution operations on the upsampled sparse feature map; Energy-aware frequency domain reconstruction normalization: Energy reconstruction operation is performed on the fused multi-band feature map. First, the local feature energy map is calculated through a set of deep convolutions, and then the input features are normalized based on the energy map. Enhanced cross-channel information interaction: Perform channel dimension rearrangement on the reconstructed feature map; Final feature fusion and output: Perform a 1×1 point convolution operation on the feature map after channel rearrangement to fuse the information of each channel in a linear combination manner.

[0009] Optionally, the multi-band deep convolution feature extraction step specifically includes: Three different convolutional kernels of 3×3, 5×5, and 7×7 are used, with the number of input channels as the number of groups, to capture low-frequency details, mid-frequency structure, and high-frequency global context features of the image respectively; The outputs of each frequency band are adaptively fused using a set of learnable channel attention weight parameters.

[0010] Optionally, the cross-channel information interaction enhancement step specifically includes: reshaping the four-dimensional feature tensor into a five-dimensional form, exchanging its second and third dimensions, and then restoring it to a four-dimensional shape.

[0011] Optionally, the illumination-sensing multi-attention fusion detection head performs the following steps in sequence: Feature input: Receives and integrates feature maps of different spatial resolutions from a multi-scale feature pyramid as input; Localization Regression: From Input Features Extract the geometric structure information of the target and predict the precise location of its bounding box; Classification confidence: from input features Extract semantic information to predict the category to which the target belongs; Illumination-sensing gating: Dynamically adjusts the intensity of classification features based on the illumination level of the features; Output fusion: Used to fuse the outputs of the localization regression step and the classification confidence step to generate detection results.

[0012] Optionally, the classification confidence step includes: Basic semantic feature extraction: Perform a 3×3 depthwise separable convolution on the input feature map to extract low-level semantic features; Feature channel dimension adjustment: The number of feature channels is adjusted by using a 1×1 standard convolutional layer and mapped to a dimension space suitable for subsequent attention processing to obtain feature G2; SEAM attention feature enhancement: First, calculate the global statistical features of each channel for feature G2, then generate channel attention weights through two fully connected layers, and finally apply the channel attention weights to feature G2. MultiSEAM attention fusion: The features processed by SEAM are fused by multiple SEAM attention heads in parallel and weighted.

[0013] Optionally, the output fusion steps include: Multi-scale feature stitching: The outputs of the localization regression step and the classification confidence step are stitched together along the channel dimension; Distribution focus regression processing: The discrete distribution parameters output by the localization regression step are converted into continuous predicted bounding box coordinates through the distribution focus regression method; Bounding box coordinate decoding: converting predicted bounding box coordinates into actual bounding box coordinates; Category probability activation: Apply the Sigmoid activation function to the classification prediction result, mapping the original output value to a probability value between 0 and 1, representing the confidence level of the target belonging to each category; Detection result generation: The decoded bounding box coordinates and the activated class probabilities are concatenated along the channel dimension to form a complete detection result output.

[0014] By adopting the above technical solution, this application has the following beneficial effects: This invention replaces the standard upsampling module in YOLOv11 with the AFRUpsample aliasing-aware frequency reconstruction upsampling module. Through an innovative frequency sensing and reconstruction mechanism, it effectively suppresses the inherent aliasing effect in the upsampling process of low-light images and significantly enhances the model's feature reconstruction capabilities for key details, small targets and blurred edges in night scenes. This invention enhances the model's ability to distinguish and extract target spectral and structural information under complex noise environments by designing an NB-SpectralFFN noise band sensing spectral feedforward module and explicitly modeling joint spatial-frequency domain feature representation, thereby improving the robustness of feature representation. This invention improves the original detection head to an LSMF-LG illumination-aware multi-attention fusion detection head. Through a dual-branch decoupling architecture and an illumination-aware gating mechanism, it optimizes the localization and classification tasks respectively. This significantly reduces the number of parameters while effectively improving the model's target recognition accuracy and bounding box regression precision under low-light conditions. In summary, this invention achieves accurate and stable detection of various nighttime targets in the ExDark dataset while ensuring inference efficiency, systematically solving key challenges such as high noise interference, easy missed detection of small targets, and weak model adaptability in low-light environments. In summary, this invention can significantly improve the detection capability and positioning accuracy of weakly visible targets, blurred edge targets, and small-scale targets while maintaining high inference speed in low-light environments. It is especially suitable for application scenarios with stringent requirements for real-time performance and accuracy, such as autonomous driving and nighttime security. Attached Figure Description

[0015] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the accompanying drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. In all the drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.

[0016] Figure 1 A flowchart of a low-light target detection method provided by an embodiment of the present invention is shown; Figure 2 The diagram shows the structure of the improved network model provided in the embodiment of the present invention; Figure 3 This is a structural diagram of the NB-SpectralFFN noise band sensing spectral feedforward neural network module provided in an embodiment of the present invention; Figure 4 A flowchart provided for embodiment of the present invention; Figure 5 A flowchart of noise perception gating provided in an embodiment of the present invention; Figure 6 This is a structural diagram of the aliasing sensing frequency reconstruction upsampling module AFRUpsample provided in an embodiment of the present invention; Figure 7 This is a structural diagram of the LSMF_LG detection head provided in an embodiment of the present invention; Figure 8 This is a structural diagram of the SEAM module provided in an embodiment of the present invention; Figure 9 This is a structural diagram of the MultiSEAM module provided in an embodiment of the present invention; Figure 10 The training results of the improved network model provided in the embodiments of the present invention are shown in the figure. Figure 11 This is a comparison of the experimental results between the improved network model and the original YOLO11 model. Detailed Implementation

[0017] The embodiments of the technical solution of the present invention will now be described in detail with reference to the accompanying drawings. These embodiments are only used to more clearly illustrate the technical solution of the present invention and are therefore merely examples, and should not be construed as limiting the scope of protection of the present invention. It should be noted that, unless otherwise stated, the technical or scientific terms used in this application should have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0018] In one embodiment, a low-light target detection method is provided, such as Figure 1 As shown, it includes: S1. The model structure is improved based on YOLOv11n. The improved model structure is as follows: Figure 2As shown, a noise frequency band sensing spectral feedforward neural network module, an aliasing sensing frequency reconstruction upsampling module, and an illumination sensing multi-attention fusion detection head are designed. Among them, the noise frequency band sensing spectral feedforward neural network module introduces a frequency domain adaptive modulation and illumination sensing dual-gating mechanism, the aliasing sensing frequency reconstruction upsampling module introduces a frequency sensing and reconstruction mechanism, and the illumination sensing multi-attention fusion detection head introduces an illumination sensing gating mechanism.

[0019] The following sections describe the three parts: the noise band sensing spectral feedforward neural network module, the aliasing sensing frequency reconstruction upsampling module, and the illumination sensing multi-attention fusion detection head.

[0020] The noise band sensing spectral feedforward neural network module NB-SpectralFFN has the following structure: Figure 3 As shown, by introducing a dual-gating mechanism of frequency domain adaptive modulation and illumination perception to explicitly model the joint feature representation of the space-frequency domain, the ability to perceive the spectral features and structural information of the target under low illumination noise environment is enhanced.

[0021] The NB-SpectralFFN processing flow includes the following stages: channel expansion and spatial context extraction, frequency domain transformation and band-aware modulation, illumination adaptive fusion, inverse transformation and noise-aware gating, feature integration and output.

[0022] A1. Channel Expansion and Spatial Context Extraction: A11. The input feature map is processed by a 1×1 convolution. The channel dimension is expanded to twice the original dimension, generating an expanded feature map. The formula used in this step is: , in, The convolution kernel weight matrix is... For bias vectors, This represents a convolution operation; the output dimension is expanded to... B is the batch size, C is the number of input channels, 2C is the number of extended channels, H is the feature map height, and W is the feature map width.

[0023] A12. Perform depthwise separable convolution on the expanded feature map to extract multi-scale spatial context information. The calculation formula is as follows: , Where b represents the batch index, c represents the channel index, i represents the height index, j represents the width index, m represents the kernel offset, and n represents the kernel offset. The weight parameters represent the depthwise separable convolution. Bias parameters for depthwise separable convolution.

[0024] A2. Frequency Domain Transformation and Band Sensing Modulation: A21. Spatial characteristics Perform a two-dimensional real-number fast Fourier transform to convert it into a frequency domain representation: , in For complex tensors, It is a two-dimensional fast Fourier transform.

[0025] A22. Construct three frequency band mask matrices based on normalized radial distance, and define the radial distance:

[0026] Where u represents the normalized vertical frequency and v represents the normalized horizontal frequency. Represents the normalized radial frequency distance. Represents the actual vertical frequency component. This represents the actual horizontal frequency component.

[0027] Generate a binary mask based on a preset threshold: , , , These correspond to the low-frequency, mid-frequency, and high-frequency bands, respectively. Indicates an indicator function.

[0028] A23. Apply learnable band weighting parameters to perform differentiated modulation on the three bands:

[0029] in The frequency domain characteristics after modulation , , For trainable parameters, This represents element-wise multiplication. Low-frequency weights enhance the main structure of the target, mid-frequency weights optimize edge details, and high-frequency weights suppress noise interference.

[0030] A3. For example Figure 4 As shown, adaptive lighting fusion: A31. Based on spatial features Generate lighting perception gating coefficients. Compress the spatial dimension using global average pooling. ;in For global average pooling, For all batches and channels at position (i,j); A32. Calculation of gating coefficients using a compressed excitation network structure:

[0031] in The weights are the weights of the first fully connected layer. For the weights of the second fully connected layer, (·) represents the Sigmoid function, which outputs g. , This is the activation function.

[0032] A33. Utilize gating coefficients to dynamically adjust the fusion ratio between the original frequency domain features and the modulated features:

[0033] in Here, g is the frequency domain bias parameter, and g is the illumination sensing gating coefficient. This mechanism achieves adaptive spectral enhancement under illumination conditions: under normal illumination (g→1), it fully utilizes band modulation, and under extremely low illumination (g→0), it preserves the original features to avoid introducing artifacts.

[0034] A4. Inverse Transform and Noise-Aware Gating: A41. Reconstruct the spatial domain representation of the processed frequency domain features using inverse Fourier transform:

[0035] Obtain reconstructed features .

[0036] A42. Divide the reconstructed features into two equal parts along the channel dimension:

[0037] in It is divided into two parts. This represents the first part of the segmentation. The second part represents the segmentation, and dim represents the segmentation along the channel dimension.

[0038] A43. Based on Calculate the noise level estimate.

[0039] First, calculate the mean of each channel:

[0040] Then calculate the variance as an estimate of the noise level: ,in For variance estimation, ) is the centered eigenvalue.

[0041] A44. Generating noise-aware gating coefficients using the Sigmoid function: , get .

[0042] A45. Combine the two features using an activation function and noise gating, such as... Figure 5 As shown, noise-sensitive feature selection is achieved: , in ; A5. Feature Integration and Output: A51. Compress the gated features back to the original channel dimension using a 1×1 convolution: , in To output the convolution kernel weights, This is the output bias.

[0043] A52. Output the final enhanced feature map This feature retains the original spectral characteristics while also possessing noise robustness and illumination adaptability, and can be directly input into subsequent network layers for further processing.

[0044] Through steps A1-A5 above, the NB-SpectralFFN module explicitly models and utilizes the multi-scale contextual information and spectral distribution characteristics of the input features in the dual-path collaborative processing of the spatial and frequency domains. Through a dual adaptive mechanism of illumination-aware gating and noise-aware gating, it dynamically adjusts the frequency domain modulation intensity and spatial feature selection weights under different lighting conditions. While enhancing the target's main structure and edge details, it effectively suppresses band noise interference and random spatial noise in low-light environments, thereby significantly improving the backbone network's noise robustness, feature representation ability, and target detection accuracy under extremely low illumination conditions.

[0045] The aliasing-sensing frequency reconstruction upsampling module AFRUpsample, such as Figure 6 As shown, through an innovative frequency sensing and reconstruction mechanism, the aliasing effect in the upsampling process is suppressed, and the robustness of the model in detecting small targets and edge features in nighttime and low-light scenes is improved.

[0046] The AFRUpsample processing flow consists of five stages: zero-value insertion upsampling operation, multi-band deep convolution feature extraction, energy-aware frequency domain reconstruction normalization, and cross-channel information interaction enhancement.

[0047] B1. Zero-value insertion upsampling operation: Receive input features Figure XA By creating an all-zero tensor and copying the original feature values ​​according to a fixed interval rule, the spatial resolution of the feature map is increased by an integer multiple, generating the output feature map Y, which is mathematically expressed as: , Where s is the upsampling factor, b is the batch index, c is the channel index, and i and j are the output spatial location indices; this non-interpolation method avoids the edge blurring and high-frequency information loss caused by traditional upsampling methods under low light conditions.

[0048] B2. Multi-band deep convolution feature extraction: The feature map output from the zero-value insertion upsampling operation is subjected to multi-scale parallel depthwise separable convolution processing, using convolution kernels of three different scales: 3×3, 5×5, and 7×7. The number of input channels is used as the number of groups, simultaneously extracting low-frequency detail features, mid-frequency structural features, and high-frequency global context features. The outputs of each convolution path are adaptively weighted and fused using a set of learnable channel attention weight parameters, the calculation formula of which is as follows: , in , , For learnable weight parameters, This indicates a channel-by-channel multiplication operation. It is a depthwise separable convolution.

[0049] B3. Energy-sensing frequency domain reconstruction normalization: An energy reconstruction operation is performed on the feature map F after multi-band deep convolution feature extraction and fusion. First, a local feature energy map E is calculated through 3×3 deep convolution. Then, the input features are normalized based on the energy map, that is, the original features are divided by the sum of the corresponding energy value and the minimum constant value, thereby suppressing noise interference in low-light images and enhancing the energy response of effective targets. The specific calculation formula is as follows: , , Where abs() is the absolute value function. It is a very small positive number.

[0050] This operation suppresses noise interference in low-light images while enhancing the energy response of effective targets.

[0051] B4. Enhanced cross-channel information exchange: The normalized feature map R reconstructed from the energy-aware frequency domain is subjected to channel-dimensional rearrangement to break the inherent inter-channel information isolation of depthwise convolution and promote the full fusion of feature information from different channels. The formula is: , , , in, To initially reshape the tensor, Let S be the tensor after dimensional permutation, S be the final output tensor, reshape be the tensor reshaping operation, and transpose be the dimensional permutation operation. This improves feature diversity without increasing the number of parameters.

[0052] B5. Final Feature Fusion and Output: A 1×1 point convolution operation is performed on the feature map S obtained through cross-channel information interaction enhancement to fuse the information from each channel in a linear combination manner, generating the final output feature map Yout. The formula is: , in, These are the point convolution weight parameters.

[0053] This step is used to output high-quality feature maps with improved resolution, enhanced semantic information, and effective noise suppression. This is used by the subsequent feature pyramid network for multi-scale feature fusion and target detection.

[0054] Through steps B1-B5 above, the aliasing-aware frequency reconstruction upsampling module AFRUpsample effectively suppresses the inherent aliasing effect in the upsampling process of low-light images through an innovative frequency perception and reconstruction mechanism. This significantly enhances the model's ability to reconstruct key details, small targets, and blurred edges in nighttime scenes, and improves the model's robustness in detecting small targets and edge features in nighttime and low-light scenes.

[0055] Light-sensing multi-attention fusion detection head, such as Figure 7 As shown, by introducing a dual-branch decoupling architecture to optimize the localization and classification tasks respectively, and by incorporating a light-sensing gating mechanism, the robustness of the model in target recognition and localization under low light conditions is explicitly enhanced.

[0056] The processing flow of the illumination-aware multi-attention fusion detection head LSMF_LG includes the following stages: feature input, localization regression, classification confidence, illumination-aware gating, and output fusion.

[0057] C1. Feature Input: Receives and integrates features from the multi-scale feature pyramid of the backbone network in the YOLOv11n baseline model.

[0058] Specifically, this unit receives three feature maps with different spatial resolutions as input, corresponding to low-resolution features in the deep layers, medium-resolution features in the mid-layers, and high-resolution features in the shallow layers, respectively. This multi-scale feature input design enables the detection head to process targets of different sizes simultaneously, laying the foundation for subsequent accurate detection.

[0059] The detection task is decomposed into independent localization regression branches and classification confidence branches.

[0060] C2. Localization Regression: Responsible for resolving input features The geometric structure information of the target is extracted, and the precise location of its bounding box is predicted. This step uses cascaded depthwise separable convolutions to construct a stable regression structure, and enhances the perception of target contours and edges through multi-layer local feature extraction.

[0061] First-level feature extraction: Extracting the input feature map... Performing a 3×3 depthwise separable convolution operation initially captures the local geometric features of the target, including basic structural information such as edges and corners. Its mathematical expression is: , in DSConv represents the input feature map, and DSConv represents depthwise separable convolution. Second-level feature enhancement: Refinement of the first-level output features Applying a third 3×3 depthwise separable convolution deepens the network's understanding of local structures, resulting in richer feature representations. Its mathematical expression is: , Third-level feature stability: for the second-level output features A third 3×3 depthwise separable convolution is performed to further stabilize the feature representation and effectively suppress the impact of noise interference on positioning accuracy. Its mathematical expression is: , Feature integration processing: The output features of the aforementioned three levels of depthwise separable convolution are integrated using a 1×1 standard convolutional layer. Information integration and fusion between channels enhances the semantic consistency of features. Its mathematical expression is: , Bounding box distribution prediction: The discrete distribution parameters of the bounding boxes are finally generated through a 1×1 convolutional layer. This layer outputs 4×R channels, where R is the regression maximum distribution value, typically set to 16. The prediction process is as follows: , in These are the regression parameters.

[0062] C3. Classification confidence: Used to evaluate input features Semantic information is extracted and enhanced through multi-level attention mechanisms to improve the discriminative power of features, ultimately predicting the category to which the target belongs. Specifically, this step uses a cascaded deep convolution (DWConv) and dual attention modules (SEAM + MultiSEAM) to achieve multi-level channel and spatial feature fusion.

[0063] Basic semantic feature extraction: First, a 3×3 depthwise convolution is performed on the input feature map to extract low-level semantic features containing target category information. Its mathematical expression is: , Feature channel dimension adjustment: The number of feature channels is adjusted using a standard 1×1 convolutional layer, mapping them to a dimension space suitable for subsequent attention processing. Its mathematical expression is: , SEAM Attention Feature Enhancement: A SEAM attention mechanism is introduced, explicitly modeling channel dependencies through compression-excitation operations between channels. The SEAM structure is as follows: Figure 8 As shown. First, calculate the global statistical characteristics for each channel: , Where c is the channel index, and H and W are the height and width of the feature map, respectively. For feature map The eigenvalue at position (i,j) in channel c.

[0064] Next, channel attention weights are generated through two fully connected layers.

[0065] First, the channel statistical features are reduced in dimensionality: , in is the dimensionality-reduced weight matrix, and ReLU is the activation function.

[0066] Secondly, the dimensionality of the reduced features is increased and the Sigmoid activation function is applied to generate channel attention weights ranging from 0 to 1: , in, For the weight matrix of increased dimension, This is the sigmoid function.

[0067] Finally, the channel attention weights are applied to the original features: , MultiSEAM Attention Layer: This layer introduces a multi-head attention mechanism on top of SEAM to model feature dependencies at multiple scales. It achieves this by parallelizing and weighted fusion of multiple SEAM attention heads. The MultiSEAM structure is as follows: Figure 9 As shown, this enhances the representational power of features: , Where K is the number of attention heads. These are learnable fusion weights.

[0068] C4. Illumination-Aware Gating: Dynamically adjusts the intensity of classification features based on the illumination level of the features. This unit performs the following operations in sequence: Global illumination statistics extraction: Perform global average pooling on the categorical features to obtain channel-level global illumination statistics. , Illumination feature compression: The number of channels is compressed to 1 / 16 of the original dimension using a 1×1 convolutional layer, reducing computational complexity while preserving key illumination information. , Illumination-gated weight generation: The feature dimension is restored to the original number of channels through a second 1×1 convolutional layer, and then normalized by the Sigmoid function to generate illumination-aware gating weights between 0 and 1. , Feature adaptive modulation: The generated illumination-aware gating weights are multiplied channel-by-channel with the original classification features to achieve adaptive enhancement and suppression of features under different illumination conditions. , Class confidence prediction: Finally, a 1×1 convolutional layer maps the modulated features to the class space, generating preliminary class confidence prediction results. , C5. Output Fusion: This function organically fuses the outputs of C2 and C3 to generate the final detection result. Specifically, the following processes are performed in sequence: Multi-scale feature stitching: combining the localization distribution parameters of three feature maps at different scales. Compared with category prediction results By concatenating the data along the channel dimension, a unified intermediate representation is formed: , Distribution focus regression processing: During the inference phase, the discrete distribution parameters output by the localization branch are converted into continuous bounding box coordinates using the distribution focus regression method.

[0069] First, the distribution parameters are converted into a probability distribution using the Softmax function: , where y is the discrete offset value and R represents the maximum value of the distribution focus regression.

[0070] Subsequently, the expected value of the bounding box offset is calculated as the continuous coordinate prediction: , In the above formula, R represents the maximum value of the distribution focus regression.

[0071] Bounding box coordinate decoding: Converting the desired offset into actual bounding box coordinates. For the bounding box center point coordinates, the calculation formula is: , , Among them, s x s represents the horizontal x-axis grid coordinates on the feature map. y d represents the vertical y-axis grid coordinates on the feature map. x d represents the continuous offset of the center point in the x-direction. y It represents the continuous offset in the y-direction from the center point. The downsampling step size of the feature map in the horizontal direction (x-axis). This represents the downsampling step size of the feature map in the vertical direction (y-axis). Category probability activation: Applying a Sigmoid activation function to the classification prediction results maps the original output values ​​to probability values ​​between 0 and 1, representing the confidence level of the target belonging to each category. , Detection result generation: The decoded bounding box coordinates and the activated class probabilities are concatenated along the channel dimension to form a complete detection result output.

[0072] Based on steps C1-C5, the illumination-aware multi-attention fusion detection head LSMF-LG optimizes the localization and classification tasks respectively through a dual-branch decoupling architecture and illumination-aware gating mechanism. While significantly reducing the number of parameters, it effectively improves the model's target recognition accuracy and bounding box regression accuracy under low illumination conditions.

[0073] S2. Train the improved network model using the ExDark low-light dataset.

[0074] The ExDark low-light dataset contains 7363 low-light images in 12 categories: bicycles, boats, bottles, buses, cars, cups, cats, dogs, motorcycles, people, and tables, all taken at night. The preprocessing steps are divided into training, validation, and test sets in an 8:1:1 ratio.

[0075] We set the training parameters, trained the model on a low-light dataset using the improved network, and accelerated the training using a GPU. The experimental environment was set to Python 3.9, CUDA 12.1, and an RTX 3080 Ti GPU for accelerated training.

[0076] The training iterations used were all 200 rounds, the batch size was set to 32, the input image size was set to 640×640, the initial learning rate was set to 0.001, the final learning rate factor was set to 0.01, the momentum parameter was set to 0.937, and the weight decay (L2 regularization coefficient) was set to 0.005.

[0077] The training model was validated by setting validation parameters. The validation results showed that the improved model achieved an improvement of 3.1% and 1.9% in mAP@0.5 and mAP@0.5:0.95 respectively compared to YOLOv11n, and the Gflops value decreased from the original 6.4 to 5.8, which achieved the requirement of improving detection accuracy while reducing computation time. After validation, the model can be loaded into the visualization interface to realize real-time detection.

[0078] S3. Use the trained model to detect low-light images.

[0079] To objectively evaluate the effectiveness of the proposed method, a systematic comparative experiment was designed in this embodiment. The comparative experiment is shown in Table 1. The performance of the improved model and the current mainstream object detection model were compared under the same low-light dataset and training configuration. Precision, recall, mean precision (mAP), number of model parameters, and computational complexity (GFLOPs) were used as evaluation indicators. Among them, precision and recall were calculated according to formulas (1) and (2) respectively, and were used to measure the accuracy and coverage of the detection results. The mean precision (mAP) was used as a comprehensive performance indicator, and its calculation was shown in formula (3), which was obtained by averaging the area under the precision-recall curves of each category. To further refine the evaluation of the model's performance under different positioning accuracy requirements, this experiment reports two results: mAP@0.5 and mAP@0.5:0.95. The former uses only the IoU threshold of 0.5 to reflect the basic detection capability; the latter takes the average mAP at multiple IoU thresholds within the range of 0.5 to 0.95 with a step size of 0.05, which can more comprehensively and rigorously evaluate the model's overall performance. Furthermore, the number of model parameters and computational complexity are used to quantitatively analyze the algorithm's efficiency and deployment feasibility.

[0080] (1) (2) (3) Where TP is the number of correctly predicted positive samples, FP is the number of incorrectly predicted positive samples, FN is the number of incorrectly predicted negative samples, AP is the area under the PR curve for the current class, and n is the number of sample classes in the dataset.

[0081] Table 1

[0082] To verify the technical contributions and synergistic effects of each improved module in this invention, ablation experiments were conducted on the ExDark low-light dataset, as shown in Table 2. Using the original YOLOv11n as a control group, the performance of individual modules and different combinations of NB-SpectralFFN (A), AFRUpsample (B), and LSMF-LG (C) were tested. Among the individual modules, A improved accuracy by 0.022, and B improved mAP@50. Modules A and C can improve recall by 0.019 and reduce computation by 1.1 GFLOPs. In the module combination, A+C achieves a balance between precision (0.728) and computation (5.7 GFLOPs), while B+C further improves mAP@50 to 0.658. After the three modules (A+B+C) of this invention work together, mAP@50 reaches 0.662 (+0.029 compared to the original model), and mAP@50-95 reaches 0.421 (+0.021 compared to the original model), while the computation decreases from 6.3 GFLOPs to 5.8 GFLOPs. Experimental results show that the modules are not simply a superposition of technologies, but rather achieve a breakthrough in both accuracy and lightweight design through a collaborative link of "spatial enhancement - feature upsampling - detection head adaptation." The experimental results of this invention are as follows: Figure 10 As shown.

[0083] Table 2

[0084] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A method for detecting targets in low light conditions, characterized in that, include: Based on YOLOv11n, the model structure was improved by designing a noise band sensing spectral feedforward neural network module, an aliasing sensing frequency reconstruction upsampling module, and an illumination sensing multi-attention fusion detection head. Among them, the noise band sensing spectral feedforward neural network module introduces a frequency domain adaptive modulation and illumination sensing dual-gating mechanism, the aliasing sensing frequency reconstruction upsampling module introduces a frequency sensing and reconstruction mechanism, and the illumination sensing multi-attention fusion detection head introduces an illumination sensing gating mechanism. The improved network model was trained using the ExDark low-light dataset. Use the trained model to detect low-light images.

2. The method according to claim 1, characterized in that, The noise frequency band sensing spectral feedforward neural network module executes the following steps in sequence: Channel expansion and spatial context extraction: The input feature map is expanded in channel dimension, and then multi-scale spatial context features are extracted; Frequency domain transformation and band-aware modulation: converting spatial domain features into frequency domain representation, constructing three frequency band mask matrices for low frequency, mid frequency, and high frequency, and performing differentiated modulation on each frequency band component through learnable frequency band weight parameters; Adaptive fusion of lighting: generating lighting perception gating coefficients based on global average pooling and compressed excitation network; Inverse Transform and Noise-Aware Gating: The frequency domain features after modulation and fusion are inversely transformed back to the spatial domain and divided into two equal parts along the channel dimension. The two parts of features are then combined through activation functions and noise gating to achieve noise-sensitive feature selection. Feature integration and output: Compress the features that have undergone noise-aware gating back to the original channel dimension.

3. The method according to claim 2, characterized in that, The aliasing-sensing frequency reconstruction upsampling module performs the following steps in sequence: Zero-value insertion upsampling operation: Performs zero-value filling upsampling based on regular grid on the input low-resolution feature map. This is achieved by creating an all-zero tensor with a spatial size that is an integer multiple of the original feature map and mapping the input feature values ​​to the corresponding grid positions of the output tensor at fixed intervals, thereby enlarging the spatial size of the feature map to the target multiple. Multi-band depthwise convolutional feature extraction: Perform multi-scale parallel depthwise separable convolution operations on the upsampled sparse feature map; Energy-aware frequency domain reconstruction normalization: Energy reconstruction operation is performed on the fused multi-band feature map. First, the local feature energy map is calculated through a set of deep convolutions, and then the input features are normalized based on the energy map. Enhanced cross-channel information interaction: Perform channel dimension rearrangement on the reconstructed feature map; Final feature fusion and output: Perform a 1×1 point convolution operation on the feature map after channel rearrangement to fuse the information of each channel in a linear combination manner.

4. The method according to claim 3, characterized in that, The multi-band deep convolution feature extraction steps specifically include: Three different convolutional kernels of 3×3, 5×5, and 7×7 are used, with the number of input channels as the number of groups, to capture low-frequency details, mid-frequency structure, and high-frequency global context features of the image respectively; The outputs of each frequency band are adaptively fused using a set of learnable channel attention weight parameters.

5. The method according to claim 3, characterized in that, The steps to enhance cross-channel information interaction specifically include: reshaping the four-dimensional feature tensor into a five-dimensional form, exchanging its second and third dimensions, and then restoring it to a four-dimensional shape.

6. The method according to claim 3, characterized in that, The illumination-sensing multi-attention fusion detection head performs the following steps in sequence: Feature input: Receives and integrates feature maps of different spatial resolutions from a multi-scale feature pyramid as input; Localization Regression: From Input Features Extract the geometric structure information of the target and predict the precise location of its bounding box; Classification confidence: from input features Extract semantic information to predict the category to which the target belongs; Illumination-sensing gating: Dynamically adjusts the intensity of classification features based on the illumination level of the features; Output fusion: Used to fuse the outputs of the localization regression step and the classification confidence step to generate detection results.

7. The method according to claim 6, characterized in that, The classification confidence step includes: Basic semantic feature extraction: Perform a 3×3 depthwise separable convolution on the input feature map to extract low-level semantic features; Feature channel dimension adjustment: The number of feature channels is adjusted by using a 1×1 standard convolutional layer and mapped to a dimension space suitable for subsequent attention processing to obtain feature G2; SEAM attention feature enhancement: First, calculate the global statistical features of each channel for feature G2, then generate channel attention weights through two fully connected layers, and finally apply the channel attention weights to feature G2. MultiSEAM attention fusion: The features processed by SEAM are fused by multiple SEAM attention heads in parallel and weighted.

8. The method according to claim 7, characterized in that, The output fusion steps include: Multi-scale feature stitching: The outputs of the localization regression step and the classification confidence step are stitched together along the channel dimension; Distribution focus regression processing: The discrete distribution parameters output by the localization regression step are converted into continuous predicted bounding box coordinates through the distribution focus regression method; Bounding box coordinate decoding: converting predicted bounding box coordinates into actual bounding box coordinates; Category probability activation: Apply the Sigmoid activation function to the classification prediction result, mapping the original output value to a probability value between 0 and 1, representing the confidence level of the target belonging to each category; Detection result generation: The decoded bounding box coordinates and the activated class probabilities are concatenated along the channel dimension to form a complete detection result output.

Images