Automobile central control screen small target detection method based on YOLOv11 improvement

Abstract

The invention discloses an automobile central control screen small target detection method based on YOLOv11 improvement, and the method specifically comprises the steps: S1, generating an image data set, carrying out the preprocessing and enhancement, and dividing the data set; s2, an improved C3k2GCConv module, a WFU module and a CGAFusion module are introduced, and a YOLOv11 network model is constructed; s3, training the model by adopting a cosine annealing learning rate and a mixed precision training strategy; s4, inputting a to-be-detected central control screen image into the improved YOLOv11 detection model, and outputting the category and bounding box coordinates of a target; and S5, performing screening and optimization through a post-processing module, and finally outputting a detection result in the form of a bounding box and a category label. According to the method, by improving the YOLOv11 model, a display target can be effectively recognized in complex environments such as strong light direct incidence, screen reflection and dim light, and the stability and robustness of the model in a complex illumination scene are improved.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence, and specifically relates to a method for detecting small targets on a car center console screen based on an improved version of YOLOv11. Background Technology

[0002] Currently, traditional object detection algorithms (such as SSD and Faster R-CNN) are mostly used to detect small objects on automotive infotainment screens. However, the application of these algorithms in automotive infotainment screen scenarios still has significant limitations. Because the display elements on infotainment screens are generally small in size and have low contrast, and the ambient lighting changes drastically (e.g., direct sunlight, reflected interference, or dim lighting), existing algorithms cannot fully preserve the detailed information and texture features of small objects during feature extraction. This leads to unstable model performance under varying lighting conditions, resulting in false positives, false negatives, and blurred edges. Furthermore, their feature fusion process does not adequately consider both high-frequency details and low-frequency structures. In complex lighting environments, target features are easily interfered with by background noise, affecting the accuracy and robustness of detection. In addition, existing methods often increase computational load while pursuing detection accuracy, leading to a decrease in inference speed and making real-time detection difficult in environments with limited computing power.

[0003] Traditional YOLO-based methods have limited feature extraction capabilities when dealing with low-brightness, small-sized display information (such as low battery indicators and small icons), easily leading to instability, missed detections, and false detections under complex lighting conditions such as direct sunlight or dim lighting. Furthermore, existing methods tend to lose high-frequency details and low-frequency structural information during neck feature fusion and detection head processing, resulting in blurred target edges and noticeable artifacts. Simultaneously, multi-scale feature fusion often involves simple weighting or stitching, failing to suppress redundant interference caused by uneven lighting or reflections, weakening key target features. When operating under limited computing power, existing technologies struggle to balance detection accuracy and real-time performance, failing to meet the high-precision, robust, and real-time requirements for small target detection on real-world control screens under complex lighting conditions.

[0004] Chinese patent document CN120510593A discloses an improved traffic sign recognition and detection method based on YOLOv11. Based on the YOLOv11 backbone network, it proposes an IPC3k2 layer, replacing standard convolutions with Ghoost to achieve lightweight feature extraction, and embedding a Coard coordinate attention mechanism in the Bottleneck to enhance spatial localization capabilities, combined with a dual-branch Ghoost feature fusion strategy. The backbone proposes a SomSPPF module, fusing multi-scale deformable pyramid pooling, a two-dimensional attention mechanism, and a SamReBlock reparameterizable convolution module. In the neck network, an MSCATR attention module is added, employing dynamic channel fusion to achieve channel statistics-guided dynamic fusion, utilizing cross-spatial fusion to establish a spatial collaboration mechanism, and introducing an adaptive feature fusion module. Finally, residual-based multi-level feature optimization is achieved through learnable weight parameters. This existing technology can effectively improve the accuracy of traffic sign detection using the YOLOv11 model, balancing the model's real-time performance and robustness. However, the model is not well adapted to dynamic scenes. It does not consider the impact of motion blur and frame rate fluctuations on the detection results during vehicle monitoring movement, and the real-time recognition stability in high-speed driving scenarios lacks verification. The model introduces multiple complex modules such as IPC3k2, SomSPPF, and MSCATR. Although the incremental increase is controlled through lightweight design, the number of parameters (3.13 million) and GFLOPs (8.2) are still increased compared to the original model, which may limit the smoothness of operation on low-computing-power embedded vehicle devices. The three-level pooling and two-dimensional attention serial processing of the SomSPPF module have certain computational redundancy, which may lead to increased inference latency.

[0005] Chinese patent document CN120278916A discloses a traffic sign dehazing detection method based on an improved MSR-YOLO, utilizing vehicle-mounted monitoring to collect various traffic sign images under foggy conditions. A multi-scale adaptive network module (MS) is used for dehazing preprocessing. The MS improves image quality and detail through three convolutional kernels of different scales, followed by dehazing through a dark channel prior dehazing filter, a gamma filter, and a sharpening filter. Based on an improved YOLOv11 network, a C3K2-RFAConv module is introduced to replace the C3K2 module in both the backbone and neck parts, and the RFAConv module in the backbone replaces the traditional convolutional module, resulting in a traffic sign detection network model, which is then trained using a dataset. The trained traffic sign detection network model is then used to perform target detection on the dehazed preprocessed image. This prior art effectively reduces the impact of foggy conditions on traffic sign recognition, thereby improving the accuracy of traffic sign detection in foggy conditions. However, the existing technology uses multi-filter cascading (dark channel prior + Gamma + sharpening) and multi-branch convolution processing, which increases inference time and makes it difficult to meet the real-time detection requirements of vehicle scenarios.

[0006] For the reasons mentioned above, traditional target detection algorithms have shortcomings in recognizing small targets on central control screens, such as low detection accuracy, poor stability, slow response speed, low detection precision, frequent false detections and missed detections, sensitivity to changes in lighting, and insufficient real-time performance. They cannot meet the comprehensive requirements of automotive central control systems for high precision and real-time performance. Summary of the Invention

[0007] The purpose of this invention is to provide a small target detection method for automotive central control screens based on an improved YOLOv11 model. This method combines multi-path feature fusion, frequency domain enhancement, and multi-dimensional attention guidance to effectively identify display targets in complex environments such as direct sunlight, screen glare, and low light conditions, improving the model's stability and robustness in complex lighting scenarios. Simultaneously, through lightweight network structure design and adaptive feature adjustment, the computational overhead during inference is reduced, achieving a balance between high accuracy and high real-time performance.

[0008] To address the aforementioned technical problems, the present invention specifically adopts the following technical solution: The method for detecting small targets on a car central control screen based on improved YOLOv11 comprises the following steps:

[0009] S1 generates an image dataset: collect images of the central control screens of several cars in traffic scenes, generate a central control screen image dataset; and preprocess and enhance the images in the image dataset, and then divide the dataset;

[0010] S2 constructs an improved YOLOv11 network model: Improved C3k2_GCConv, WFU and CGAFusion modules are introduced to improve the original YOLOv11 backbone network and feature fusion structure, and construct a YOLOv11 network model.

[0011] S3 Model Training: The preprocessed and enhanced central control screen image dataset is input into the improved YOLOv11 network for training. During the training process, a cosine annealing learning rate and mixed precision training strategy are adopted to dynamically optimize the model parameters, so as to accelerate network convergence and improve the stability and generalization ability of feature extraction. After training, the improved YOLOv11 detection model is obtained.

[0012] S4 Model Inference: The image of the central control screen to be detected is input into the trained improved YOLOv11 detection model. First, multi-scale feature maps are extracted through the backbone network. After being fused by the improved C3k2_GCConv module, multi-scale feature maps are generated. Then, feature fusion is performed by the WFU module, followed by feature optimization by the CGAFusion module. Finally, the target category and bounding box coordinates are output. Through the fusion prediction of multi-scale features, feature amplification and redundancy suppression of salient regions of small targets are achieved, hierarchical feature reconstruction and context enhancement are realized, and adaptive recognition and localization output of targets of different sizes are achieved.

[0013] S5 Detection Result Post-processing: The set of predicted bounding boxes output by the model is filtered and optimized by the post-processing module, and the detection results are finally output in the form of bounding boxes and category labels.

[0014] By employing the aforementioned technical solutions and combining a YOLOv11 improved structure with multi-path feature fusion, frequency domain enhancement, and multi-dimensional attention guidance, detection stability and small target recognition rate can be significantly improved under complex lighting and low contrast conditions. A C3k2_GCConv module is introduced into the backbone network to achieve efficient extraction and representation of small target features through multi-branch convolutional fusion and structural reparameterization mechanisms, maintaining the stability of feature representation under complex lighting and texture interference conditions, thereby reducing the false detection rate of small targets. A WFU module is introduced in the neck region to enhance high-frequency details and low-frequency structures through wavelet decomposition and inverse transform, reducing information loss and artifact effects, improving the edge fidelity of small targets under low light conditions, and mitigating false detection problems. A CGAFusion module is set in front of the detection head, utilizing channel, spatial, and pixel-level attention to achieve adaptive feature fusion, effectively suppressing interference caused by uneven lighting or reflection, highlighting key target information, and thus improving overall detection accuracy. This method achieves high-precision detection of small targets on the central control screen under limited computing power and maintains stronger robustness and real-time performance under complex lighting conditions.

[0015] Preferably, in step S1, all images in the central control screen image dataset are labeled with corresponding category labels and bounding box coordinates; and multiple enhancement methods are used for preprocessing, specifically: multiple images are randomly stitched together in the same frame using the Mosaic enhancement strategy to simulate the distribution of targets at different scales; at the same time, the images are randomly flipped and the brightness is adjusted to enhance the model's adaptability to lighting, blur and noise; the enhanced dataset is divided into a training set (about 70%), a validation set (about 20%) and a test set (about 10%) to ensure that the model obtains sufficient feature learning during the training phase and has effective generalization verification.

[0016] Preferably, in step S2, the internal structure of the introduced C3k2_GCConv module includes three main components: the GCConv module, the Bottleneck_GCConv module, and the C3k_GCConv module, which are nested in a hierarchical relationship.

[0017] First, the input features are enhanced locally using the Bottleneck_GCConv module;

[0018] Subsequently, the features are fed into the multi-branch structure of the GCConv module. Multiple convolutional branches first extract spatial detail information under different receptive fields, and then another convolutional branch performs channel compression and mapping enhancement. At the same time, a residual signal path is provided to pass the original features. The outputs of the multiple branches are fused element-wise in the channel dimension and non-linearly mapped by the SiLU activation function to obtain the fused enhanced features.

[0019] It contains two Convolutional branch, one Convolution branch and one residual signal path; two The convolutional branches are used to extract spatial detail information under different receptive fields. The convolutional branch is used for channel compression and mapping enhancement, while the residual signal path directly transmits the original features to ensure gradient stability when the input and output channels are the same. The outputs of the four paths are fused element-wise in the channel dimension and nonlinearly mapped by the SiLU activation function to obtain the enhanced features after fusion.

[0020] During the training phase, the GCConv module enhances the model's feature representation ability and gradient flow stability through a multi-path convolutional structure.

[0021] During the inference phase, multi-path convolution and batch normalization parameters are integrated into a single parameter through parameter fusion and convolution kernel reparameterization. Convolution kernel;

[0022] The final output feature map contains both spatial details and global semantic information.

[0023] The specific steps for feature extraction by the C3k2_GCConv module in step S2 are as follows:

[0024] At the lowest level, the GCConv module employs three parallel paths to achieve multi-scale contextual fusion of features; input features via two Convolutional branch (denoted as) ) and a Convolutional branches Features are extracted in parallel; additionally, if in == out and stride == 1, a batch normalized residual path is introduced. The three features are weighted and summed and activated during the forward propagation process. The processed output is:

[0025] ;

[0026] During the training phase, the GCConv module enhances the model's feature representation ability and gradient flow stability through a multi-path convolutional structure.

[0027] During the inference phase, the three-path convolution is transformed into a single equivalent 3×3 convolution kernel through parameter fusion, thereby achieving structural re-parameterization and ensuring computational efficiency and model consistency.

[0028] The Bottleneck_GCConv module, built on the GCConv module, replaces the original two-layer standard convolution with a two-layer GCConv stacked structure in the traditional Bottleneck structure.

[0029] Let the input channel be The output channel is The Bottleneck_GCConv module is then represented as:

[0030] ;

[0031] Where α∈{0,1} indicates whether shortcut residual connections are enabled; when shortcut is True, the module can enhance the nonlinear expressive power of features without increasing computational complexity; This indicates the second GCConv layer (which further extracts semantic features and restores the channel dimension). This indicates the first GCConv layer (used for channel compression or preliminary feature extraction), and X represents the input feature tensor of the module (i.e., the input feature map).

[0032] In higher-level structures, the C3k_GCConv module uses Bottleneck_GCConv as its core stacking unit, repeating it n times to form a sequence, expressed as:

[0033] ;

[0034] After the module input is split, one path extracts deep contextual features through the sequence module, while the other path retains the original feature flow. The two feature paths are concatenated and fused with a 1×1 convolution to form a composite feature representation that takes into account both fine-grained and global semantics.

[0035] M represents the main branch structure consisting of n Bottleneck_GCConv modules stacked sequentially; Bottleneck_GCConvᵢ represents the i-th Bottleneck_GCConv unit, where i is the stack index variable (i=1,…,n), used to extract feature representations at different levels sequentially.

[0036] Ultimately, the C3k2_GCConv module implements two different structural forms based on the Boolean value of the parameter c3k: when c3k=False, the main branch consists of multiple layers of Bottleneck_GCConv; when c3k=True, it consists of multiple layers of stacked C3k_GCConv; its overall output characteristic Y is represented as:

[0037] ;

[0038] in, The main branch is a stacked module, where X1 and X2 are the input features after channel segmentation; Conv represents the convolution operation; and Concat represents the concatenation operation.

[0039] Through this design, C3k2_GCConv structurally balances the fine-grained capture of local textures with comprehensive modeling of the global context, significantly improving the robustness and accuracy of small object detection while maintaining moderate computational complexity. Based on a multi-path reparameterizable structure, the C3k2_GCConv module achieves multi-scale feature fusion through three-path parallel convolution and residual connections. It enhances feature representation capabilities during training and simplifies the structure through parameter fusion during inference, thus balancing accuracy and speed. This design effectively improves the robustness of small object representation on the central control screen under complex lighting conditions and reduces missed detections.

[0040] Preferably, in step S2, the WFU module introduces wavelet decomposition and inverse transform mechanisms during the YOLOv11 feature fusion stage.

[0041] First, the WFU module receives feature inputs from different scales, including high-resolution feature maps (x_big) and low-resolution semantic feature maps (x_small) output from the backbone network. During the input stage, the WFU module performs two-dimensional discrete wavelet decomposition on the high-resolution features using the built-in HaarWavelet wavelet transform, dividing the original features into four sub-bands: the low-frequency component a (Approximation) is used to characterize global brightness and structural information, while the high-frequency components h (Horizontal), v (Vertical), and d (Diagonal) characterize the detailed features in the horizontal, vertical, and diagonal directions, respectively.

[0042] Then, the high-frequency components h, v, d are summed element-wise and input into the high-frequency fusion feature extraction module (RefinementBlock, RB). The high-frequency fusion feature extraction module consists of two 3×3 convolution layers, which are used to refine the edge and texture response and enhance the salient features under reflected light and noise conditions, so as to strengthen the model's response to the edge area of ​​the small icons on the central control screen.

[0043] Meanwhile, the low-frequency component a is concatenated with the low-resolution semantic feature map x_small from the shallow layer in the channel dimension, and the dimension mapping and information interaction are performed through the channel transformation network, so as to inject low-level semantic information into the low-frequency structure to achieve cross-scale compensation.

[0044] Finally, the high-frequency fusion features processed by the high-frequency fusion feature extraction module and the low-frequency compensated features obtained by channel transformation are concatenated by channel, and then reconstructed back into the spatial domain through inverse Haar wavelet transform, achieving joint spatial-frequency reconstruction of the features. The final output is a fusion feature map with the same resolution as the input. This reconstruction step achieves synergistic optimization of the frequency and spatial domains, providing robust input with both detail and semantics for subsequent detection heads. By utilizing wavelet feature upgrading and adaptive fusion strategies, the feature representation ability and generalization performance of the model are effectively improved while maintaining detection speed.

[0045] The WFU module extracts features by taking two feature maps from the backbone network as input. ,in Indicates high-resolution features, It represents low-resolution features; the WFU module structure consists of three parts: a wavelet feature decomposition layer, a frequency domain enhancement and channel transform layer, and a wavelet inverse transform fusion layer;

[0046] First, the wavelet decomposition layer (HaarWavelet) processes the input features. Performing a two-dimensional Haar wavelet transform yields four subbands: a low-frequency component A, a horizontal high-frequency component H, a vertical high-frequency component V, and a diagonal high-frequency component D, which can be mathematically expressed as follows:

[0047] ;

[0048] The calculation for each sub-band can be expressed as:

[0049] ;

[0050] ;

[0051] in, Indicates the feature map in the local The pixel blocks within the window, together with the four sub-bands, constitute the feature group after scale decomposition.

[0052] Secondly, the frequency domain enhancement layer extracts high-frequency fusion features through blocks. Nonlinear fusion and edge texture enhancement of high-frequency components h, v, and d are performed to obtain frequency domain enhanced features. The formula is:

[0053] ;

[0054] Simultaneously, the low-frequency component A is combined with low-resolution features from shallow layers. Channel splicing and feature transformation are performed to obtain the enhanced low-frequency compensation features. The formula is:

[0055] ;

[0056] in, This represents a Channel Transformation network consisting of two convolutional layers; 'a' represents the low-frequency component.

[0057] Finally, the fusion layer reconstructs the enhanced high and low frequency features using inverse Haar wavelet transform, outputting the reconstructed feature Out, expressed as:

[0058] ;

[0059] Here, IHaar represents the inverse wavelet transform operation, which is the inverse of the forward wavelet decomposition, and realizes the joint spatial-frequency reconstruction of features.

[0060] The WFU module can simultaneously retain the fine-grained boundary information of shallow features and the global semantic expression of deep features, thereby achieving the distribution and balance of feature energy in target detection at different scales. It is particularly suitable for the stable recognition of reflective, small-sized targets such as central control screens. By introducing the WFU module, high- and low-frequency feature fusion is achieved through Haar wavelet decomposition and inverse reconstruction, enhancing the model's detail perception ability and feature expression stability, thus significantly improving the robustness of small target detection under complex lighting and background conditions. The WFU module is used to achieve adaptive feature fusion with channel, spatial, and pixel-level attention in front of the detection head, thereby effectively suppressing light reflection, background interference, and redundant noise information during multi-scale feature integration, highlighting the salient features of the target region.

[0061] Preferably, in step S2, the CGAFusion module takes two sets of input features (x and y) from different feature layers as input. First, it forms an initial fusion feature map by adding them element-wise to unify the feature space and semantic scale. Then, the module sequentially calculates the weight mappings of the three sub-modules: channel attention, spatial attention, and pixel attention, achieving multi-level feature modulation from global to local levels. Specifically:

[0062] In the channel attention stage, the CGAFusion module first extracts the global response features of each channel through adaptive average pooling, and then passes them through two layers. Convolution and ReLU activation form the channel weight distribution (CATTN), which is used to characterize the importance of different channels, thereby strengthening the response to key semantic feature channels and suppressing redundant or background channels.

[0063] In the Spatial Attention stage, the CGAFusion module first performs average pooling and max pooling operations on the input feature maps along the channel dimension to obtain the channel average feature map and the channel maximum feature map; then, it concatenates the two along the channel dimension to form a dual-channel feature representation, and then... Convolution generates a spatial attention weight map (SATTN); this weight map is used to highlight salient target regions in the spatial dimension and suppress background interference, thereby enhancing the spatial responsiveness of small target features.

[0064] Building upon the first two attention mechanisms, the CGAFusion module further introduces pixel-level attention for fine-grained feature calibration; the pixel attention module combines the initial feature map with the comprehensive attention map obtained in the previous stage ( =sattn+cattn) is fused, and a pixel-wise weight distribution map is generated through channel rearrangement and depthwise separable convolution. The pixel layer is dynamically adjusted through Sigmoid activation. This process enables the model to adaptively allocate weights on each pixel, refine the feature discrimination ability, and thus effectively avoid feature drift or occlusion of small objects in complex backgrounds.

[0065] Finally, the CGAFusion module, guided by pixel attention weights, performs a weighted fusion of the two sets of input features, calculated as follows:

[0066] ;

[0067] Where x and y are two sets of input features, respectively; This is the comprehensive attention map from the previous stage; =sattn+cattn, where sattn is the spatial attention weight map and cattn is the channel weight distribution; It is based on the fusion of features (x+y) and comprehensive attention. The generated pixel-level attention map is used to dynamically adjust the fusion ratio of x and y at each location;

[0068] and through Convolution further integrates and fuses the results, outputting a feature map with significant target response and interference suppression information.

[0069] The CGAFusion module enables explicit association of multi-scale features and pixel-level semantic enhancement without significantly increasing computational cost, effectively improving YOLOv11's ability to recognize small targets on central control screens in complex backgrounds. The CGAFusion module is designed to hierarchically fuse spatial attention, channel attention, and pixel attention, achieving adaptive weighting and dynamic information selection of multimodal features, effectively enhancing the model's feature representation capabilities and adaptability to complex scenes.

[0070] The CGAFusion module is used for dynamic guidance and weighted reconstruction of cross-scale features during the YOLOv11 feature fusion stage. Through a cascaded fusion mechanism of channel attention, spatial attention, and pixel attention, it achieves information interaction and selective enhancement between multi-scale features, thereby improving the detection sensitivity of small targets and weakly textured regions. Specifically:

[0071] The input consists of two feature maps from different scales. Where B represents the batch size, C represents the number of channels, H represents the height of the feature map, and W represents the width of the feature map; first, element-wise addition is performed to obtain the initial fused features. The expression is:

[0072] ;

[0073] This initial fusion serves as input to channel attention and spatial attention to extract global dependency and local response information;

[0074] The ChannelAttention (CGA) branch extracts the global channel descriptor using adaptive average pooling, expressed as:

[0075] ;

[0076] Where c represents the channel index; i, j represent the spatial coordinates of the feature map; H, W represent the height and width of the feature map, respectively; This represents the global average eigenvalue of the c-th channel;

[0077] Then through two layers Convolution and ReLU activation form channel attention weights The expression is:

[0078] ;

[0079] in, , These are the convolution weight matrices, For ReLU functions;

[0080] The SpatialAttention_CGA branch generates two spatial descriptors through max pooling and average pooling along the channel dimension, expressed as:

[0081] ;

[0082] in, This indicates the average pooling operation. This represents the feature map obtained from the average pooling operation. This indicates a max pooling operation. This represents the feature map obtained from the max pooling operation;

[0083] Then, splicing and convolution operations are performed to obtain a spatial attention map. The expression is:

[0084] ;

[0085] in, Indicates reflective fill Convolution operation;

[0086] Subsequently, the two attention-weighted features are fused to obtain the first-stage pixel-guided attention input features. The expression is:

[0087] ;

[0088] Among them, the pixel attention branch (PixelAttention_CGA) uses input features Features of the first stage By concatenating along the new dimension, we obtain the new dimension concatenated feature X, expressed as:

[0089] ;

[0090] After dimensional rearrangement, rearranged features are obtained. The expression is:

[0091] ;

[0092] And pixel-level attention maps are generated using depthwise separable convolution groups=C:

[0093] ;

[0094] in, This represents the Sigmoid function, and the graph is used to guide the weighting of two input features at the pixel level.

[0095] The final fusion result Out is represented as:

[0096] ;

[0097] in, Indicates reflective fill Convolution operation; Adaptive weight allocation is performed on the feature channels of input x and y to achieve pixel-level guided fusion of cross-scale features.

[0098] Through this design, the CGAFusion module can achieve explicit association of multi-scale features and pixel-level semantic enhancement without significantly increasing the amount of computation, which can effectively enhance YOLOv11's ability to recognize small targets on the central control screen in complex backgrounds.

[0099] Preferably, the loss function used when training the model in step S3 includes bounding box regression loss. Confidence loss and category loss Its overall loss function is expressed as:

[0100] ;

[0101] in, These are the weighting coefficients used to control the weight distribution between detection accuracy and positioning accuracy.

[0102] Bounding box regression loss The formula is expressed as:

[0103] ;

[0104] Where B=(x,y,w,h) represents the center coordinates and width and height of the prediction box; The center coordinates and width and height of the ground truth bounding box; gt refers to the ground truth bounding box, which is used to compare with the model's predicted bounding box to calculate the loss. This represents the intersection-union ratio (IoU) between the predicted bounding box and the ground truth bounding box. represents the squared Euclidean distance between the centers of the predicted bounding box and the ground truth bounding box; c represents the diagonal length of the bounding boxes of the predicted bounding box and the ground truth bounding box. Indicates the weighting adjustment factor; This indicates a measure of the difference in aspect ratio.

[0105] Confidence loss The formula is expressed as:

[0106] ;

[0107] in, This indicates that the image is divided into S×S grid boxes; Indicates whether the i-th grid / anchor box has a target (1 indicates yes, 0 indicates no); This represents the probability that the model predicts there is a target in the i-th box.

[0108] Category loss The formula is expressed as:

[0109] ;

[0110] in, This represents the number of predicted boxes that actually contain the target; C represents the total number of categories. This is represented as 1 when the i-th box actually belongs to the c-th class, and 0 otherwise; This represents the probability that the model predicts the i-th bounding box belongs to the c-th class. Preferably, in step S3, a cosine annealing learning rate adjustment strategy is used during model training to achieve dynamic optimization of the learning rate, and stochastic gradient descent (SGD) optimization algorithm is combined to balance the convergence speed and stability of the model; simultaneously, mixed precision training (FP16) and gradient accumulation are introduced. While ensuring controllable GPU memory usage, more efficient model training and parameter updates are achieved, thereby improving the stability and robustness of the improved YOLOv11 detection model in small object recognition tasks.

[0111] Preferably, the model inference in step S4 specifically involves:

[0112] Let the input image be After network mapping, the predicted tensor is output. ,in Represents the bounding box coordinates. For category confidence, The probability of the target existing; the final detection score is calculated using the following formula:

[0113] .

[0114] Preferably, the post-processing module performs the following filtering and optimization steps in step S5:

[0115] S51 Confidence Filtering: Setting a confidence threshold Filter by confidence level, retain only The candidate box;

[0116] S52 Non-Maximum Suppression (NMS) Redundancy Removal: An improved NMS algorithm is used to remove redundancy from overlapping targets. When the Intersection over Union (IoU) between two candidate boxes exceeds a threshold... At that time, only the player with the higher score will be retained;

[0117] S53 Multi-scale Fusion Correction: For target areas of the central control screen with large scale differences, a weighted fusion strategy between feature pyramid layers is introduced to compensate for displacement errors of small targets and improve detection integrity.

[0118] Significant scale differences refer to the fact that the same target (such as the central control screen) appears in different sizes in the image due to changes in distance, angle, or resolution under different shooting conditions. The final output detection results are given in the form of bounding boxes and category labels to ensure high detection robustness even in complex lighting and reflection environments.

[0119] Preferably, the method further includes step S6: model deployment and application. After model training, the improved YOLOv11 detection model is deployed to the actual system for detection application. After deployment, the improved YOLOv11 detection model receives real-time image input from the central control screen, automatically completes feature extraction, target recognition and localization, and finally outputs the detection result after post-processing operations such as confidence filtering, non-maximum suppression (NMS), and multi-scale fusion correction. The detection accuracy on the automotive central control screen dataset is significantly improved compared to the original YOLOv11 model, especially in small target detection tasks where the false negative rate is significantly reduced. This algorithm can be widely applied to scenarios such as automotive central control display status recognition, intelligent cockpit monitoring, and vehicle anomaly alerts.

[0120] Compared with the prior art, the technical solution of the present invention has the following beneficial technical effects:

[0121] (1) Enhance small target detection capability: The C3k2_GCConv module is introduced to realize multi-path convolution feature fusion, enabling the model to aggregate multi-scale semantic and texture information at the same time during the feature extraction stage, which significantly improves the detection accuracy and stability of small targets on the central control screen.

[0122] (2) Enhance multi-scale information fusion: By using the WFU module to perform wavelet decomposition, frequency domain enhancement and inverse transformation reconstruction on high and low resolution features, dynamic compensation and channel interaction are achieved, reducing information loss and improving the blurring of small target edges and loss of details.

[0123] (3) Enhance feature fusion and discrimination capabilities: The CGAFusion module achieves adaptive fusion of different feature domains through the synergistic effect of multi-dimensional attention mechanism, which improves the saliency and feature contrast of the target region and makes the detection results more accurate and stable.

[0124] (4) Maintaining high-efficiency inference performance: The improved network maintains detection accuracy while having a lightweight structure and a limited increase in the number of parameters. Its inference speed is similar to that of the original YOLOv11 model, making it suitable for embedded devices and vehicle computing platforms.

[0125] (5) The model has a simple design structure, low computational load, good engineering applicability and scalability, and is suitable for deployment in actual intelligent cockpit and central control screen interaction systems to realize real-time target detection function. Attached Figure Description

[0126] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0127] Figure 1 This is a flowchart of the small target detection method for automotive central control screens based on YOLOv11, as described in this invention.

[0128] Figure 2 This is a structural diagram of the C3k2_GCConv module in the YOLOv11-based improved method for detecting small targets on a car's central control screen, where C3k2 is in a closed or disabled state.

[0129] Figure 3 This is a structural diagram of the C3k2_GCConv module of the small target detection method for automotive central control screens based on YOLOv11, as described in this invention.

[0130] Figure 4 This is a structural diagram of the WFU module in the improved small target detection method for automotive central control screens based on YOLOv11 of this invention;

[0131] Figure 5 This is a structural diagram of the CGAFusion module of the small target detection method for automotive central control screens based on YOLOv11, as described in this invention.

[0132] Figure 6 This is a diagram showing the overall network structure of the algorithm for the small target detection method for automotive central control screens based on YOLOv11, as described in this invention.

[0133] Figure 7 A graph showing the changes in model training and validation loss and performance metrics;

[0134] Figure 8 This is a structural diagram of the existing YOLOv11 model. Detailed Implementation

[0135] The embodiments of the present invention will be described in detail below with reference to the accompanying drawings. The following embodiments are only used to illustrate the technical solutions of the present invention more clearly, and should not be used to limit the scope of protection of the present invention.

[0136] Those skilled in the art will understand that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. It should also be understood that terms such as those defined in general dictionaries should be understood to have the same meaning as in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless defined as herein.

[0137] Common knowledge in this invention includes the Faster R-CNN algorithm and the YOLOv11 object detection algorithm;

[0138] Faster R-CNN is a classic two-stage object detection algorithm and one of the fundamental and well-known techniques for small object detection. Its core structure includes a backbone network, a Region Proposal Network (RPN), and a head. In the workflow, the input image first passes through the convolutional backbone network to extract multi-level features. These features contain both low-level spatial information and high-level semantic information about the target, providing a foundation for candidate region generation and classification. Subsequently, the Region Proposal Network uses a sliding window to generate candidate boxes and combines them with feature maps to predict the target's location and size, providing candidate regions (RoIs) for the head. The head then performs classification and bounding box regression on each candidate box, achieving accurate target localization and recognition.

[0139] The Faster R-CNN algorithm is characterized by high accuracy, especially in the detection of dense and small targets. However, its two-stage detection process results in a slower detection speed, making it unsuitable for scenarios with high real-time requirements. Furthermore, for small targets or scenes with significant lighting variations, the candidate regions generated by the RPN may miss detections, and insufficient feature fusion may lead to the loss or blurring of target edge information. These limitations provide important reference and comparison for single-stage target detection algorithms and the improvements of this invention, namely, how to improve speed and real-time performance while maintaining detection accuracy.

[0140] The YOLOv11 object detection algorithm is a single-stage object detection algorithm in the YOLO series, featuring end-to-end fast detection capabilities and optimizations in small object detection and multi-scale feature utilization. Its network structure mainly consists of a backbone, a neck, and a head. The backbone extracts multi-level image features through convolutional layers and residual modules. The neck uses a Feature Pyramid Network (FPN) structure for multi-scale feature fusion, enabling low-level small object information to be effectively transmitted to higher levels for recognition, thus enhancing small object detection capabilities. The head directly outputs predicted bounding boxes and category information, and uses improved loss functions (such as GIoU and CIoU) to optimize the regression accuracy of the predicted bounding boxes. YOLOv11 introduces data augmentation methods such as Mosaic and MixUp during the training phase, enriching training samples through random combination, scaling, and cropping of images, improving the model's robustness under complex lighting, occlusion, or reflection conditions. Furthermore, YOLOv11 optimizes network structure and computational efficiency, enabling real-time detection even in environments with limited computing power.

[0141] The advantages of the YOLOv11 object detection algorithm in small object detection, fast inference, and multi-scale feature fusion provide the foundation for this invention. This allows the invention to further introduce C3k2_GCConv, WFU, and CGAFusion modules to enhance the small object feature extraction capability, edge fidelity, and illumination robustness of the central control screen, thereby achieving high-precision, robust, and real-time small object detection.

[0142] Example: The specific steps of this improved YOLOv11-based method for detecting small targets on a car's central control screen are as follows:

[0143] S1 generates an image dataset: collect images of the central control screens of several cars in traffic scenes, generate a central control screen image dataset; and preprocess and enhance the images in the image dataset, and then divide the dataset;

[0144] In step S1, the images in the central control screen image dataset are labeled with corresponding category labels and bounding box coordinates; and multiple enhancement methods are used for preprocessing. The Mosaic enhancement strategy is used to randomly stitch multiple images into the same screen to simulate the distribution of targets at different scales; at the same time, the images are randomly flipped and the brightness is adjusted to enhance the model's adaptability to lighting, blur and noise; the dataset is divided into training set, validation set and test set.

[0145] S2 constructs an improved YOLOv11 network model: Improved C3k2_GCConv, WFU, and CGAFusion modules are introduced to enhance the original YOLOv11 backbone network and feature fusion structure, thus constructing the YOLOv11 network model. Specifically, the C3k2_GCConv module is introduced into the backbone to enhance feature extraction and cross-channel information interaction capabilities; the WFU module is introduced into the neck to achieve multi-scale feature weighted fusion; and the CGAFusion module is added to the multi-scale output to enhance the interaction and integration of features at different levels.

[0146] In step S2, the C3k2_GCConv module is a lightweight feature extraction structure proposed in this invention, used to enhance the feature representation capability of small targets on the central control screen under complex lighting conditions. This module introduces multi-branch convolution and structural reparameterization mechanisms based on the traditional C3 structure. Through the Bottleneck_GCConv substructure and parallel GCConv branches, it achieves efficient extraction and fusion of multi-scale features, thereby enhancing the model's robustness to lighting changes and its ability to capture details while maintaining computational efficiency.

[0147] like Figure 2 and Figure 3 As shown, the internal structure of the improved C3k2_GCConv module includes three main components: the GCConv module, the Bottleneck_GCConv module, and the C3k_GCConv module. The GCConv module, the Bottleneck_GCConv module, and the C3k_GCConv module are nested in a hierarchical relationship.

[0148] First, the input features are enhanced locally using the Bottleneck_GCConv module;

[0149] Subsequently, the features are fed into the multi-branch structure of the GCConv module. Multiple convolutional branches first extract spatial detail information under different receptive fields, and then another convolutional branch performs channel compression and mapping enhancement. At the same time, a residual signal path is provided to pass the original features. The outputs of the multiple branches are fused element-wise in the channel dimension and non-linearly mapped by the SiLU activation function to obtain the fused enhanced features.

[0150] It contains two Convolutional branch, one Convolution branch and one residual signal path; two The convolutional branches are used to extract spatial detail information under different receptive fields. The convolutional branch is used for channel compression and mapping enhancement, while the residual signal path directly transmits the original features to ensure gradient stability when the input and output channels are the same. The outputs of the four paths are fused element-wise in the channel dimension and nonlinearly mapped by the SiLU activation function to obtain the enhanced features after fusion.

[0151] During the training phase, the GCConv module enhances the model's feature representation ability and gradient flow stability through a multi-path convolutional structure.

[0152] During the inference phase, multi-path convolution and batch normalization parameters are integrated into a single parameter through parameter fusion and convolution kernel reparameterization. Convolutional kernels effectively reduce model computation and improve inference speed.

[0153] The final output feature map combines spatial details with global semantic information. It can maintain stable feature responses in complex environments such as strong light reflection and low light, providing high-quality input for subsequent feature enhancement and attention fusion modules.

[0154] A WFU (Wavelet Feature Upgrade) module is introduced into the neck structure of the network. This module utilizes wavelet decomposition and inverse transform to decompose the input features into high-frequency details and low-frequency structure. The high-frequency component is used to enhance the target edge contours, while the low-frequency component preserves global semantic features. Weighted fusion is then used to maintain the integrity of the image structure, thereby improving edge fidelity under low-light conditions.

[0155] In step S2, the WFU module introduces wavelet decomposition and inverse transform mechanisms in the YOLOv11 feature fusion stage to simultaneously enhance high-frequency details and low-frequency structural information, thereby improving the model's ability to detect small targets under complex lighting conditions. The overall process of this module is as follows:

[0156] First, the WFU module receives feature inputs from different scales, including high-resolution feature maps (x_big) and low-resolution semantic feature maps (x_small) output from the backbone network. During the input stage, the WFU module performs two-dimensional discrete wavelet decomposition on the high-resolution features using the built-in HaarWavelet wavelet transform, dividing the original features into four sub-bands: the low-frequency component a (Approximation) is used to characterize global brightness and structural information, while the high-frequency components h (Horizontal), v (Vertical), and d (Diagonal) characterize the detailed features in the horizontal, vertical, and diagonal directions, respectively.

[0157] Then, the high-frequency components h, v, d are summed element-wise and input into the high-frequency fusion feature extraction module (RefinementBlock, RB). The high-frequency fusion feature extraction module consists of two 3×3 convolution layers, which are used to refine the edge and texture response and enhance the salient features under reflected light and noise conditions, so as to strengthen the model's response to the edge area of ​​the small icons on the central control screen.

[0158] Meanwhile, the low-frequency component a is concatenated with the low-resolution semantic feature map x_small from the shallow layer in the channel dimension, and the dimension mapping and information interaction are performed through the channel transformation network, so as to inject low-level semantic information into the low-frequency structure to achieve cross-scale compensation.

[0159] Finally, the high-frequency fusion features processed by the high-frequency fusion feature extraction module RB and the low-frequency compensation features obtained by channel transformation are concatenated by channel, and reconstructed back into the spatial domain through inverse Haar wavelet transform, achieving joint spatial-frequency feature restoration. The final output is a fusion feature map with the same resolution as the input. This reconstruction step achieves synergistic optimization of the frequency and spatial domains, providing robust input with both detail and semantics for subsequent detection heads. By utilizing wavelet feature upgrading and adaptive fusion strategies, the feature representation ability and generalization performance of the model are effectively improved while maintaining detection speed. Through this series of decomposition, enhancement, and reconstruction processes, the WFU module achieves synergistic optimization of the frequency and spatial domains at the feature level, enabling the model to maintain clear edges, stable structure, and sufficient feature representation for small targets even in complex lighting scenarios such as direct sunlight, screen reflection, or dim lighting. This effectively alleviates false detection and false negative problems and provides more robust feature input for subsequent detection heads. The WFU module structure diagram is shown below. Figure 4 As shown;

[0160] A CGAFusion module (multi-level attention fusion module) is set in front of the detection head. Through a three-level feature weighting mechanism of channel attention, spatial attention and pixel attention, the importance of features in different regions is adaptively adjusted, effectively suppressing false feature responses caused by uneven illumination and reflection interference, thereby highlighting the key information of the target area and improving the stability and accuracy of the detection results.

[0161] In step S2, the CGAFusion module takes two sets of input features (x and y) from different feature layers as input. First, it forms an initial fusion feature map by adding them element-wise to unify the feature space and semantic scale. Then, the module sequentially calculates the weight mappings of the three sub-modules: channel attention, spatial attention, and pixel attention, achieving multi-level feature modulation from global to local levels. Specifically:

[0162] In the channel attention stage, the CGAFusion module first extracts the global response features of each channel through adaptive average pooling, and then passes them through two layers. Convolution and ReLU activation form the channel weight distribution (CATTN), which is used to characterize the importance of different channels, thereby strengthening the response to key semantic feature channels and suppressing redundant or background channels.

[0163] In the Spatial Attention stage, the CGAFusion module first performs average pooling and max pooling operations on the input feature maps along the channel dimension to obtain the channel average feature map and the channel maximum feature map; then, it concatenates the two along the channel dimension to form a dual-channel feature representation, and then... Convolution generates a spatial attention weight map (SATTN); this weight map is used to highlight salient target regions in the spatial dimension and suppress background interference, thereby enhancing the spatial responsiveness of small target features.

[0164] Building upon the first two attention mechanisms, the CGAFusion module further introduces pixel-level attention for fine-grained feature calibration; the pixel attention module combines the initial feature map with the comprehensive attention map obtained in the previous stage ( =sattn+cattn) is fused, and a pixel-wise weight distribution map is generated through channel rearrangement and depthwise separable convolution. The pixel layer is dynamically adjusted through Sigmoid activation. This process enables the model to adaptively allocate weights on each pixel, refine the feature discrimination ability, and thus effectively avoid feature drift or occlusion of small objects in complex backgrounds.

[0165] Finally, the CGAFusion module, guided by pixel attention weights, performs a weighted fusion of the two sets of input features, calculated as follows:

[0166] ;

[0167] Where x and y are two sets of input features, respectively; This is the comprehensive attention map from the previous stage; =sattn+cattn, where sattn is the spatial attention weight map and cattn is the channel weight distribution; It is based on the fusion of features (x+y) and comprehensive attention. The generated pixel-level attention map is used to dynamically adjust the fusion ratio of x and y at each location. And through... The convolution further integrates and fuses the results, outputting a feature map with significant target response and suppression of interference information. This module achieves collaborative optimization of the channel layer, spatial layer, and pixel layer, effectively highlighting small target areas and suppressing false features caused by reflection interference and uneven brightness in complex lighting environments of automotive central control screens. This provides the detection head with more discriminative and stable input features, significantly improving detection accuracy and robustness.

[0168] The CGAFusion module enables explicit association of multi-scale features and pixel-level semantic enhancement without significantly increasing computational cost, effectively improving YOLOv11's ability to recognize small targets on central control screens in complex backgrounds. The CGAFusion module is designed to hierarchically fuse spatial attention, channel attention, and pixel attention, achieving adaptive weighting and dynamic information selection of multimodal features, effectively enhancing the model's feature representation capabilities and adaptability to complex scenes. The CGAFusion module structure diagram is shown below. Figure 5 As shown;

[0169] S3 Model Training: The preprocessed and enhanced central control screen image dataset is input into the improved YOLOv11 network for training. During the training process, a cosine annealing learning rate and mixed precision training strategy are adopted to dynamically optimize the model parameters, so as to accelerate network convergence and improve the stability and generalization ability of feature extraction. After training, the improved YOLOv11 detection model is obtained.

[0170] The loss function used in step S3 when training the model includes bounding box regression loss. Confidence loss and category loss Its overall loss function is expressed as:

[0171] ;

[0172] in, These are the weighting coefficients for each item, used to control the weight distribution between detection accuracy and positioning accuracy;

[0173] Bounding box regression loss The formula is expressed as:

[0174] ;

[0175] Where B=(x,y,w,h) represents the center coordinates and width and height of the prediction box; The center coordinates and width and height of the ground truth bounding box; gt refers to the ground truth bounding box, which is used to compare with the model's predicted bounding box to calculate the loss. This represents the intersection-union ratio (IoU) between the predicted bounding box and the ground truth bounding box. represents the squared Euclidean distance between the centers of the predicted bounding box and the ground truth bounding box; c represents the diagonal length of the bounding boxes of the predicted bounding box and the ground truth bounding box. Indicates the weighting adjustment factor; This indicates a measure of the difference in aspect ratio;

[0176] Confidence loss The formula is expressed as:

[0177] ;

[0178] in, This indicates that the image is divided into S×S grid boxes; Indicates whether the i-th grid / anchor box has a target (1 indicates yes, 0 indicates no); This represents the probability that the model predicts there is a target in the i-th box.

[0179] Category loss The formula is expressed as:

[0180] ;

[0181] in, This represents the number of predicted boxes that actually contain the target; C represents the total number of categories. This is represented as 1 when the i-th box actually belongs to the c-th class, and 0 otherwise; This represents the probability that the model predicts the i-th box belongs to the c-th class;

[0182] In step S3, a cosine annealing learning rate adjustment strategy is used to dynamically optimize the learning rate during model training, and stochastic gradient descent (SGD) is combined to balance the model's convergence speed and stability. Simultaneously, mixed precision training (FP16) and gradient accumulation are introduced. While ensuring controllable GPU memory usage, this achieves more efficient model training and parameter updates, thereby improving the stability and robustness of the improved YOLOv11 detection model in small target recognition tasks.

[0183] S4 Model Inference: The image of the central control screen to be detected is input into the trained improved YOLOv11 detection model. First, multi-scale feature maps are extracted through the backbone network. After being fused by the improved C3k2_GCConv module, multi-scale feature maps are generated. Then, feature fusion is performed by the WFU module, followed by feature optimization by the CGAFusion module. Finally, the target category and bounding box coordinates are output. Through the fusion prediction of multi-scale features, feature amplification and redundancy suppression of salient regions of small targets are achieved, hierarchical feature reconstruction and context enhancement are realized, and adaptive recognition and localization output of targets of different sizes are achieved.

[0184] In step S4, the model inference process specifically involves:

[0185] Let the input image be After network mapping, the predicted tensor is output. ,in Represents the bounding box coordinates. For category confidence, The probability of the target existing; the final detection score is calculated using the following formula:

[0186] ;

[0187] S5 Detection Result Post-processing: The set of predicted bounding boxes output by the model is filtered and optimized by the post-processing module, and the detection results are finally output in the form of bounding boxes and category labels;

[0188] The specific steps for filtering and optimization performed by the post-processing module in step S5 are as follows:

[0189] S51 Confidence Filtering: Setting a confidence threshold Filter by confidence level, retain only The candidate box;

[0190] S52 Non-Maximum Suppression (NMS) Redundancy Removal: An improved NMS algorithm is used to remove redundancy from overlapping targets. When the Intersection over Union (IoU) between two candidate boxes exceeds a threshold... At that time, only the player with the higher score will be retained;

[0191] S53 Multi-Scale Fusion Correction: For central control screen target areas with large scale differences (large scale differences refer to the same target (such as the central control screen) appearing in different sizes in the image due to changes in distance, angle or resolution under different shooting conditions, and the final output detection results are given in the form of bounding boxes and category labels to ensure high detection robustness in complex lighting and reflection environments), a feature pyramid layer weighted fusion strategy is introduced to compensate for small target displacement errors and improve detection completeness.

[0192] The final output detection results are given in the form of bounding boxes and category labels to ensure high detection robustness even in complex lighting and reflection environments;

[0193] S6 Model Deployment and Application: After the model training is completed, the improved YOLOv11 detection model is deployed to the actual system for detection application; after deployment, the improved YOLOv11 detection model receives real-time image input from the central control screen, automatically completes feature extraction, target recognition and localization, and finally outputs the detection result after post-processing operations such as confidence filtering, non-maximum suppression (NMS) and multi-scale fusion correction.

[0194] The overall structure of this algorithm consists of three parts: Backbone (main feature extraction layer), Neck (feature fusion layer), and Head (detection layer). This structure is specifically optimized based on YOLOv11, introducing modules such as C3k2_GCConv, WFU (Wavelet Feature Upgrade), and CGAFusion to enhance multi-scale feature representation capabilities and improve small object detection performance. Figure 6 As shown.

[0195] The backbone is primarily responsible for multi-level feature extraction from the input image. The input image first undergoes two layers of Conv convolutional operations for preliminary feature encoding to capture low-level spatial texture information. Subsequently, the features are passed through multiple levels of C3k2_GCConv modules to achieve efficient interaction between features and suppression of redundant features, thereby extracting discriminative mid-to-high-level semantic information. In the high-level feature extraction stage, the SPPF (Spatial Pyramid Pooling-Fast) module expands the network's receptive field, enhancing the model's ability to perceive multi-scale targets. Then, the C2PSA module is used to fuse channel attention and spatial attention mechanisms, adaptively strengthening key semantic feature responses and suppressing background interference, providing more robust high-level semantic information for subsequent feature fusion stages.

[0196] The Neck section is used to achieve multi-level fusion and enhancement of features at different scales. First, upsampling and weighted feature fusion are achieved through two WFU (Wavelet Feature Upgrade) modules. Wavelet decomposition is used to preserve high-frequency details during feature reconstruction, reducing information loss due to scale differences between feature layers. The fused features are further refined using the C3k2_GCConv module to generate more discriminative multi-scale feature maps. Simultaneously, the Neck section employs Concat and Conv operations in multiple places to achieve horizontal and vertical information interaction, enabling the full fusion of shallow detail features and deep semantic features, enhancing the network's robustness to small targets and complex backgrounds. Furthermore, CGAFusion modules are deployed at all three outputs of the Neck. This module combines channel attention and spatial attention mechanisms to adaptively aggregate feature information from different levels before outputting to the detector head, strengthening the response of salient target regions, suppressing redundant features, and achieving deep fusion of cross-scale and cross-layer semantics.

[0197] The Head section contains three detection branches at different scales, each receiving features from the Neck at different scales, corresponding to the detection tasks of small, medium, and large targets, and performing the final category prediction and bounding box regression.

[0198] Specific application examples:

[0199] S1 generates a central control screen image dataset: A large number of images of car central control screens are collected through in-vehicle cameras or public data sources, ensuring data coverage of various typical scenarios, including complex environments such as direct sunlight, low-light nighttime conditions, screen glare, and dust obstruction. Small target areas within the central control screen, such as battery icons and brightness warnings, are highlighted in the collected images. Each sample is labeled with its corresponding category label and bounding box coordinates, enabling the model to learn the spatial distribution patterns of multiple categories of small targets during training. To improve the model's generalization performance and robustness, various enhancement operations are performed on the original dataset. A Mosaic enhancement strategy is used to randomly stitch multiple images together in the same frame, simulating target distribution at different scales; random flipping and brightness adjustment are also used to enhance the model's adaptability to lighting, blur, and noise. The enhanced dataset is proportionally divided into a training set (approximately 70%), a validation set (approximately 20%), and a test set (approximately 10%) to ensure the model acquires sufficient feature learning during training and provides effective generalization validation.

[0200] S2 constructs an improved YOLOv11 network structure: Based on the original YOLOv11 backbone network and feature fusion structure, the following three innovative modules are introduced to improve the performance of feature extraction, semantic fusion and small target recognition.

[0201] The specific steps for feature extraction by the C3k2_GCConv module in step S2 are as follows:

[0202] At the lowest level, the GCConv module employs three parallel paths to achieve multi-scale contextual fusion of features; input features via two Convolutional branch (denoted as) ) and a Convolutional branches Features are extracted in parallel; additionally, if in == out and stride == 1, a batch normalized residual path is introduced. The three features are weighted and summed and activated during the forward propagation process. The processed output is:

[0203] ;

[0204] During the training phase, the GCConv module enhances the model's feature representation ability and gradient flow stability through a multi-path convolutional structure.

[0205] During the inference phase, the three-path convolution is transformed into a single equivalent 3×3 convolution kernel through parameter fusion, thereby achieving structural re-parameterization and ensuring computational efficiency and model consistency.

[0206] The Bottleneck_GCConv module, built on the GCConv module, replaces the original two-layer standard convolution with a two-layer GCConv stacked structure in the traditional Bottleneck structure.

[0207] Let the input channel be The output channel is The Bottleneck_GCConv module is then represented as:

[0208] ;

[0209] Where α∈{0,1} indicates whether shortcut residual connections are enabled; when shortcut is True, the module can enhance the nonlinear expressive power of features without increasing computational complexity; This indicates the second GCConv layer (which further extracts semantic features and restores the channel dimension). This indicates the first GCConv layer (used for channel compression or preliminary feature extraction), and X represents the input feature tensor of the module (i.e., the input feature map).

[0210] In higher-level structures, the C3k_GCConv module uses Bottleneck_GCConv as its core stacking unit, repeating it n times to form a sequence, expressed as:

[0211] ;

[0212] After the module input is split, one path extracts deep contextual features through the sequence module, while the other path retains the original feature flow. The two feature paths are concatenated and fused with a 1×1 convolution to form a composite feature representation that takes into account both fine-grained and global semantics.

[0213] M represents the main branch structure consisting of n Bottleneck_GCConv modules stacked sequentially; Bottleneck_GCConvᵢ represents the i-th Bottleneck_GCConv unit, where i is the stack index variable (i=1,…,n), used to extract feature representations at different levels sequentially.

[0214] Ultimately, the C3k2_GCConv module implements two different structural forms based on the Boolean value of the parameter c3k: when c3k=False, the main branch consists of multiple layers of Bottleneck_GCConv; when c3k=True, it consists of multiple layers of stacked C3k_GCConv; its overall output characteristic Y is represented as:

[0215] ;

[0216] in, The main branch is a stacked module, where X1 and X2 are the input features after channel segmentation; Conv represents the convolution operation; and Concat represents the concatenation operation.

[0217] Through this design, C3k2_GCConv structurally balances the fine-grained capture of local textures with comprehensive modeling of the global context, significantly improving the robustness and accuracy of small object detection while maintaining moderate computational complexity. Based on a multi-path reparameterizable structure, the C3k2_GCConv module achieves multi-scale feature fusion through three-path parallel convolution and residual connections. It enhances feature representation capabilities during training and simplifies the structure through parameter fusion during inference, thus balancing accuracy and speed. This design effectively improves the robustness of small object representation on the central control screen under complex lighting conditions and reduces missed detections.

[0218] The WFU module extracts features by taking two feature maps from the backbone network as input. ,in Indicates high-resolution features, It represents low-resolution features; the WFU module structure consists of three parts: a wavelet feature decomposition layer, a frequency domain enhancement and channel transform layer, and a wavelet inverse transform fusion layer;

[0219] First, the wavelet decomposition layer (HaarWavelet) processes the input features. Performing a two-dimensional Haar wavelet transform yields four subbands: a low-frequency component A, a horizontal high-frequency component H, a vertical high-frequency component V, and a diagonal high-frequency component D, which can be mathematically expressed as follows:

[0220] ;

[0221] The calculation for each sub-band can be expressed as:

[0222] ;

[0223] ;

[0224] in, Indicates the feature map in the local The pixel blocks within the window, together with the four sub-bands, constitute the feature group after scale decomposition.

[0225] Secondly, the frequency domain enhancement layer extracts high-frequency fusion features through blocks. Nonlinear fusion and edge texture enhancement of high-frequency components h, v, and d are performed to obtain frequency domain enhanced features. The formula is:

[0226] ;

[0227] Simultaneously, the low-frequency component A is combined with low-resolution features from shallow layers. Channel splicing and feature transformation are performed to obtain the enhanced low-frequency compensation features. The formula is:

[0228] ;

[0229] in, This represents a Channel Transformation network consisting of two convolutional layers; 'a' represents the low-frequency component.

[0230] Finally, the fusion layer reconstructs the enhanced high and low frequency features using inverse Haar wavelet transform, outputting the reconstructed feature Out, expressed as:

[0231] ;

[0232] Here, IHaar represents the inverse wavelet transform operation, which is the inverse of the forward wavelet decomposition, and realizes the joint spatial-frequency reconstruction of features.

[0233] The WFU module can simultaneously preserve the fine-grained boundary information of shallow features and the global semantic representation of deep features, thereby achieving the distribution and balance of feature energy in target detection at different scales. It is particularly suitable for the stable recognition of reflective, small-sized targets such as central control screens. By introducing the WFU module, high- and low-frequency feature fusion is achieved through Haar wavelet decomposition and inverse reconstruction, enhancing the model's detail perception ability and feature representation stability, thus significantly improving the robustness of small target detection under complex lighting and background conditions.

[0234] The CGAFusion module is used for dynamic guidance and weighted reconstruction of cross-scale features during the YOLOv11 feature fusion stage. Through a cascaded fusion mechanism of channel attention, spatial attention, and pixel attention, it achieves information interaction and selective enhancement between multi-scale features, thereby improving the detection sensitivity of small targets and weakly textured regions. Specifically:

[0235] The input consists of two feature maps from different scales. Where B represents the batch size, C represents the number of channels, H represents the height of the feature map, and W represents the width of the feature map; first, element-wise addition is performed to obtain the initial fused features. The expression is:

[0236] ;

[0237] This initial fusion serves as input to channel attention and spatial attention to extract global dependency and local response information;

[0238] The ChannelAttention (CGA) branch extracts the global channel descriptor using adaptive average pooling, expressed as:

[0239] ;

[0240] Where c represents the channel index; i, j represent the spatial coordinates of the feature map; H, W represent the height and width of the feature map, respectively; This represents the global average eigenvalue of the c-th channel;

[0241] Then through two layers Convolution and ReLU activation form channel attention weights The expression is:

[0242] ;

[0243] in, , These are the convolution weight matrices, For ReLU functions;

[0244] The SpatialAttention_CGA branch generates two spatial descriptors through max pooling and average pooling along the channel dimension, expressed as:

[0245] ;

[0246] in, This indicates the average pooling operation. This represents the feature map obtained from the average pooling operation. This indicates a max pooling operation. This represents the feature map obtained from the max pooling operation;

[0247] Then, splicing and convolution operations are performed to obtain a spatial attention map. The expression is:

[0248] ;

[0249] in, Indicates reflective fill Convolution operation;

[0250] Subsequently, the two attention-weighted features are fused to obtain the first-stage pixel-guided attention input features. The expression is:

[0251] ;

[0252] Among them, the pixel attention branch (PixelAttention_CGA) uses input features Features of the first stage By concatenating along the new dimension, we obtain the new dimension concatenated feature X, expressed as:

[0253] ;

[0254] After dimensional rearrangement, rearranged features are obtained. The expression is:

[0255] ;

[0256] And pixel-level attention maps are generated using depthwise separable convolution groups=C:

[0257] ;

[0258] in, This represents the Sigmoid function, and the graph is used to guide the weighting of two input features at the pixel level.

[0259] The final fusion result Out is represented as:

[0260] ;

[0261] in, Indicates reflective fill Convolution operation; Adaptive weight allocation is performed on the feature channels of input x and y to achieve pixel-level guided fusion of cross-scale features.

[0262] Through this design, the CGAFusion module can achieve explicit association of multi-scale features and pixel-level semantic enhancement without significantly increasing the amount of computation, which can effectively enhance YOLOv11's ability to recognize small targets on the central control screen in complex backgrounds.

[0263] S3 Input Training Data and Model Training: The preprocessed training set is input into the improved YOLOv11 model for training; the SGD optimizer is used for parameter updates, the learning rate is initially set to 0.01, and a cosine annealing strategy is used for dynamic adjustment; the loss function consists of classification loss, localization loss, and confidence loss; during training, enhancement strategies such as Mosaic and MixUp are used to improve the robustness of the model, and finally the trained model weights are obtained;

[0264] The loss function typically consists of three parts for object detection: bounding box regression loss. Confidence loss With category loss The overall loss function is expressed as:

[0265] ;

[0266] in, These are the weighting coefficients for each item, used to control the weight distribution between detection accuracy and positioning accuracy;

[0267] After the model training is completed, the improved YOLOv11 network is deployed to a real system for detection applications. The model has good versatility and lightweight characteristics, and can run stably on GPU servers, embedded devices, or in-vehicle terminals.

[0268] S4 Model Inference Phase: In the inference phase, the image of the central control screen to be detected is input into the trained improved YOLOv11 model. The model first extracts multi-scale feature maps through the backbone network, and then performs hierarchical feature reconstruction and context enhancement through the improved C3k2_GCConv module, WFU module and CGAFusion module to achieve feature amplification and redundancy suppression of salient regions of small targets;

[0269] Let the input image be After network mapping, the predicted tensor is output. ,in Represents the bounding box coordinates. For category confidence, The probability of the target existing. The final detection score is calculated using the following formula:

[0270]

[0271] The model achieves adaptive recognition and localization output of targets of different sizes by fusing and predicting multi-scale features;

[0272] S5 Detection Result Post-processing: The set of predicted bounding boxes output by the model is filtered and optimized by the post-processing module. This mainly includes confidence filtering: setting a confidence threshold. Only retain Candidate boxes; Non-maximum suppression (NMS): An improved NMS algorithm is used to remove redundancy from overlapping targets. When the intersection-over-union (IoU) ratio between two candidate boxes exceeds a threshold... When the target area has a large scale difference (meaning the same target (e.g., the central control screen) appears in different sizes in the image due to changes in distance, angle, or resolution under different shooting conditions, the final output detection results are given in the form of bounding boxes and class labels to ensure high detection robustness under complex lighting and reflection environments), a feature pyramid layer weighted fusion strategy is introduced to compensate for small target displacement errors and improve detection completeness. The final output detection results are given in the form of bounding boxes and class labels to ensure high detection robustness under complex lighting and reflection environments.

[0273] S6 Model Deployment and Application: After model training is complete, the improved YOLOv11 network is deployed to a real-world system for detection applications. The model features good versatility and lightweight characteristics, and can run stably on GPU servers, embedded devices, or in-vehicle terminals, allowing for visualization of detection results.

[0274] The ablation experiments were conducted on the original YOLOv11 detection model and the improved model (YOLOv11n+C3k2_GCConv+WFU+CGAFusion) of this invention, and the specific results are shown in Table 1.

[0275] Table 1 Comparison Results of Ablation Experiments

[0276] Evaluation index formula:

[0277] The formula for precision (P) is as follows:

[0278] ;

[0279] Where TP represents the number of correctly predicted positive samples; FP represents the number of incorrectly predicted positive samples.

[0280] The recall rate (R) is expressed as follows:

[0281] ;

[0282] Wherein, FN represents the number of true targets that were missed.

[0283] ;

[0284] Where C represents the total number of categories; This represents the average precision of class c; IoU=0.5 indicates that the threshold for classifying a predicted bounding box as correct (TP) is 0.5.

[0285] ;

[0286] The IoU threshold ranges from 0.5 to 0.95, with a step size of 0.05 (a total of 10 thresholds). The AP of each category is calculated first, and then the average is taken.

[0287] Table 1 shows that the mAP0.5 and mAP0.5:0.95 of the YOLOv11n model are 83.5% and 56.0%, respectively. When only the C3k2_GCConv module is introduced, the model's mAP0.5 increases to 85.0%, indicating that multi-branch convolution and reparameterized structures effectively enhance feature extraction capabilities. When the WFU module is introduced alone, the mAP0.5 further increases to 85.2%, indicating that the fusion of high- and low-frequency features based on wavelet decomposition can improve detail perception and structure preservation capabilities. When the CGAFusion module is introduced alone, the mAP0.5 reaches 85.1%, indicating that the multi-dimensional attention fusion mechanism achieves effective feature collaboration at the channel, spatial, and pixel levels. Figure 7 As shown, the first row represents the training set metrics, specifically:

[0288] train / box_loss represents the bounding box regression loss during the training phase. The lower the loss, the more accurate the model's prediction of the target box location. As shown in the figure, the loss decreases rapidly with each training iteration, indicating that the model is learning the target location.

[0289] train / cls_loss represents the classification loss during the training phase, which measures the accuracy of the model's prediction of the target class. The lower the value, the stronger the model's ability to distinguish between different classes.

[0290] train / dfl_loss represents the distributed bounding box loss during the training phase. The lower the value of the distributed bounding box loss during the training phase, the more accurate the bounding box prediction distribution.

[0291] Metrics / precision(B) represents the precision during the training phase, which indicates how many samples predicted as positive are actually positive (TP / (TP+FP)). A higher value indicates better detection performance.

[0292] metrics / recall(B) represents the recall rate during the training phase. Recall indicates how many real positive samples are correctly predicted by the model (TP / (TP+FN)). A higher value indicates that most targets are detected.

[0293] The second row represents the validation set metrics, specifically:

[0294] val / box_loss represents the bounding box regression loss on the validation set, used to test the model's generalization ability. The downward trend in the graph is similar to that on the training set, indicating that there is no obvious overfitting.

[0295] val / cls_loss represents the classification loss on the validation set, and as above, it is used to evaluate the generalization ability of the model in class prediction.

[0296] val / dfl_loss represents the distributed bounding box loss on the validation set; the lower the value, the better the bounding box prediction performance on the validation set.

[0297] metrics / mAP50(B) represents the mAP0.5 of the validation set (IoU threshold is 0.5), which is a comprehensive indicator of detection accuracy. The higher the value, the better.

[0298] metrics / mAP50-95(B) represents the mAP of the validation set as 0.5:0.95 (average of IoU from 0.5 to 0.95). Considering the detection accuracy under different IoU thresholds, a higher value is better.

[0299] When C3k2_GCConv, WFU, and CGAFusion are used in combination, the model performance reaches its optimal level, with mAP0.5 increasing to 87.4% and mAP0.5:0.95 increasing to 59.3%, representing improvements of 3.9% and 3.3% respectively compared to the baseline model. The results indicate that each module has significant complementarity in feature extraction, detail enhancement, and feature fusion, and their combined use can achieve high-precision detection of small targets on the central control screen.

[0300] The improved YOLOv11 detection model of the present invention is compared with the model in the prior art, and the specific results are shown in Table 2.

[0301] Table 2 Comparison results between the improved YOLOv11 detection model and existing models.

[0302]

[0303] As shown in Table 2, the improved YOLOv11 model outperforms the original YOLOv11n and YOLOv8n algorithms in all performance metrics. Specifically, precision is improved by approximately 2.9% compared to YOLOv11n, recall by 1.8%, mAP0.5 by 3.9%, and mAP0.5:0.95 by 3.3%. Furthermore, the model's inference speed is increased to 100.4, approximately 35.7% faster than YOLOv8n.

[0304] The results show that the present invention significantly improves the accuracy and real-time performance of small target detection while maintaining the network's lightweight nature, fully verifying the effectiveness and synergistic enhancement of the proposed C3k2_GCConv, WFU, and CGAFusion modules.

[0305] This invention primarily utilizes the YOLOv11 network as the detection framework. During the feature extraction and fusion stages, it employs multi-branch convolutional fusion and structural reparameterization mechanisms to efficiently extract and represent features at different scales, thereby improving the detection stability of small targets under complex lighting conditions. Simultaneously, by combining wavelet domain feature enhancement and multi-level attention fusion methods, it enhances the model's adaptability to lighting changes while preserving edge details and texture features. This structure balances detection accuracy and computational efficiency, providing a more robust foundation for small target recognition on automotive central control screens under complex lighting conditions.

[0306] The main problem addressed by this invention is the shortcomings of traditional object detection algorithms in central control screen applications, including low detection accuracy, frequent false positives and false negatives, sensitivity to changes in lighting, and insufficient real-time performance. By introducing feature enhancement and attention fusion mechanisms, this invention can effectively identify display targets in environments with strong direct sunlight, screen reflections, and low light, improving the model's stability and robustness in complex lighting scenarios. Simultaneously, through lightweight network structure design and adaptive feature adjustment, the computational overhead during inference is reduced, achieving a balance between high accuracy and high real-time performance.

[0307] To address the shortcomings of conventional deep learning algorithms in detection, this invention overcomes the problems of insufficient response to high-frequency details and unstable feature representation in small target recognition by improving the network feature extraction and fusion path. The improved structure strengthens multi-scale feature association, enabling the network to focus more on key detection regions and avoid feature drift caused by illumination reflection or background interference, thereby improving the accuracy and consistency of detection.

[0308] The algorithm's core process revolves around "feature extraction—detail enhancement—adaptive fusion." First, the network employs a multi-branch convolutional structure to capture hierarchical features, jointly modeling the semantic and detailed information of targets at different scales. Then, a collaborative optimization strategy in the frequency and spatial domains enhances edge and texture features, ensuring the target retains complete feature representation even under strong or weak light conditions. Finally, an adaptive fusion mechanism integrates channel, spatial, and pixel-level features, dynamically allocating weights to key region information. This overall process constructs a highly efficient and robust detection system, enabling accurate identification and real-time detection of small targets on the central control screen even in complex lighting environments.

[0309] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.

Claims

1. A method for detecting small targets on a car central control screen based on an improved version of YOLOv11, characterized in that, Specifically, the following steps are included: S1 generates an image dataset: collect images of the central control screens of several cars in traffic scenes, generate a central control screen image dataset; and preprocess and enhance the images in the image dataset, and then divide the dataset; S2 constructs an improved YOLOv11 network model: Improved C3k2_GCConv, WFU and CGAFusion modules are introduced to improve the original YOLOv11 backbone network and feature fusion structure, and construct a YOLOv11 network model. S3 Model Training: The preprocessed and enhanced central control screen image dataset is input into the improved YOLOv11 network for training. During the training process, a cosine annealing learning rate and mixed precision training strategy are adopted to dynamically optimize the model parameters, so as to accelerate network convergence and improve the stability and generalization ability of feature extraction. After training, the improved YOLOv11 detection model is obtained. S4 Model Inference: The image of the central control screen to be detected is input into the trained improved YOLOv11 detection model. First, multi-scale feature maps are extracted through the backbone network. After being fused by the improved C3k2_GCConv module, multi-scale feature maps are generated. Then, feature fusion is performed by the WFU module, and feature optimization is performed by the CGAFusion module. Finally, the target category and bounding box coordinates are output. S5 Detection Result Post-processing: The set of predicted bounding boxes output by the model is filtered and optimized by the post-processing module, and the detection results are finally output in the form of bounding boxes and category labels.

2. The method for detecting small targets on a car central control screen based on YOLOv11 as described in claim 1, characterized in that, In step S1, the images in the central control screen image dataset are labeled with corresponding category labels and bounding box coordinates; and multiple enhancement methods are used for preprocessing, specifically: the Mosaic enhancement strategy is used to randomly stitch multiple images into the same screen to simulate the distribution of targets at different scales; at the same time, the images are randomly flipped and the brightness is adjusted, and the enhanced dataset is divided into training set, validation set and test set according to the proportion.

3. The method for detecting small targets on a car central control screen based on YOLOv11 as described in claim 1, characterized in that, In step S2, the internal structure of the improved C3k2_GCConv module includes: GCConv module, Bottleneck_GCConv module and C3k_GCConv module, which are nested in a hierarchical relationship. First, the input features are enhanced locally using the Bottleneck_GCConv module; Subsequently, the features are fed into the multi-branch structure of the GCConv module. Multiple convolutional branches first extract spatial detail information under different receptive fields, and then another convolutional branch performs channel compression and mapping enhancement. At the same time, a residual signal path is provided to pass the original features. The outputs of the multiple branches are fused element-wise in the channel dimension and non-linearly mapped by the SiLU activation function to obtain the fused enhanced features. During the training phase, the GCConv module enhances the model's feature representation ability and gradient flow stability through a multi-path convolutional structure. During the inference phase, multi-path convolution and batch normalization parameters are integrated into a single convolution kernel through parameter fusion and kernel reparameterization operations. The final output feature map contains both spatial details and global semantic information.

4. The method for detecting small targets on a car central control screen based on YOLOv11 as described in claim 3, characterized in that, In step S2, the improved WFU module introduces wavelet decomposition and inverse transform mechanisms in the YOLOv11 feature fusion stage. First, the WFU module receives feature inputs from different scales, including a high-resolution feature map x_big and a low-resolution semantic feature map x_small output from the backbone network. During the input stage, the WFU module performs two-dimensional discrete wavelet decomposition on the high-resolution features using the built-in HaarWavelet wavelet transform, dividing the original features into four sub-bands: the low-frequency component a is used to characterize global brightness and structural information, while the high-frequency components h, v, and d characterize the detailed features in the horizontal, vertical, and diagonal directions, respectively. Then, the high-frequency components h, v, and d are summed element-wise and input into the high-frequency fusion feature extraction module; the high-frequency fusion feature extraction module consists of two 3×3 convolutional layers. Meanwhile, the low-frequency component a is concatenated with the low-resolution semantic feature map x_small from the shallow layer in the channel dimension, and the dimension mapping and information interaction are performed through the channel transformation network, so as to inject low-level semantic information into the low-frequency structure to achieve cross-scale compensation. Finally, the high-frequency fusion features processed by the high-frequency fusion feature extraction module and the low-frequency compensation features obtained by channel transformation are concatenated by channel, and reconstructed back into the spatial domain by inverse wavelet transform to achieve joint spatial-frequency restoration of features. The final output is a fusion feature map with the same resolution as the input.

5. The method for detecting small targets on a car central control screen based on YOLOv11 as described in claim 4, characterized in that, In step S2, the improved CGAFusion module takes two sets of input features from different feature layers as input. First, it forms an initial fusion feature map by adding elements one by one, which is used to unify the feature space and semantic scale. Then, the module calculates the weight mapping of the three sub-modules of channel attention, spatial attention and pixel attention in sequence to realize multi-level feature modulation from global to local. Specifically: During the channel attention stage, the CGAFusion module first extracts the global response features of each channel through adaptive average pooling, and then processes them through two layers. Convolution and ReLU activation form the channel weight distribution; In the spatial attention stage, the CGAFusion module first performs average pooling and max pooling operations on the input feature map along the channel dimension to obtain the channel average feature map and the channel max feature map; then, the two are concatenated along the channel dimension to form a dual-channel feature representation, and then... Convolution generates a spatial attention weight map; Based on the first two attention mechanisms, the CGAFusion module introduces a pixel attention module, which fuses the initial feature map with the comprehensive attention map obtained in the previous stage, generates a pixel-wise weight distribution map through channel rearrangement and depthwise separable convolution, and realizes dynamic adjustment of the pixel layer through Sigmoid activation. Finally, the CGAFusion module, guided by pixel attention weights, performs a weighted fusion of the two sets of input features, calculated as follows: ； Where x and y are two sets of input features; This is the comprehensive attention map from the previous stage; =sattn+cattn, where sattn is the spatial attention weight map and cattn is the channel weight distribution; It is based on the fusion of features (x+y) and comprehensive attention. The generated pixel-level attention map is used to dynamically adjust the fusion ratio of x and y at each location; Finally, pass through The convolution process further integrates and fuses the results, outputting a fused feature map.

6. The method for detecting small targets on a car central control screen based on YOLOv11 as described in any one of claims 3-5, characterized in that, The loss function used in step S3 when training the model includes bounding box regression loss. Confidence loss and category loss Its overall loss function is expressed as: ； in, These are the weighting coefficients used to control the weight distribution between detection accuracy and positioning accuracy.

7. The method for detecting small targets on a car central control screen based on YOLOv11 as described in claim 6, characterized in that, In step S3, a cosine annealing learning rate adjustment strategy is used to dynamically optimize the learning rate during model training, and the stochastic gradient descent (SGD) optimization algorithm is combined to balance the convergence speed and stability of the model; at the same time, mixed precision training and gradient accumulation are introduced.

8. The method for detecting small targets on a car central control screen based on YOLOv11 as described in claim 7, characterized in that, In step S4, the model inference process specifically involves: Let the input image be After network mapping, the predicted tensor is output. ,in Represents the bounding box coordinates, where Indicates the coordinates of the center point of the bounding box; This represents the width of the bounding box; This represents the height of the bounding box; For category confidence, The probability of the target existing; the final detection score is calculated using the following formula: 。 9. The method for detecting small targets on a car central control screen based on YOLOv11 as described in claim 7, characterized in that, The specific steps for filtering and optimization performed by the post-processing module in step S5 are as follows: S51 Confidence Filtering: Setting a confidence threshold Filter by confidence level, retain only The candidate box; S52 Non-Maximum Suppression Redundancy Removal: An improved NMS algorithm is used to remove redundancy from overlapping targets. When the Intersection over Union (IoU) between two candidate boxes exceeds a threshold... At that time, only the player with the higher score will be retained; S53 Multi-scale Fusion Correction: For target areas of the central control screen with large scale differences, a weighted fusion strategy between feature pyramid layers is introduced to compensate for displacement errors of small targets and improve detection integrity.

10. The method for detecting small targets on a car central control screen based on YOLOv11 as described in claim 7, characterized in that, It also includes step S6, model deployment and application: After the model training is completed, the improved YOLOv11 detection model is deployed to the actual system for detection application; after deployment, the improved YOLOv11 detection model receives real-time image input from the central control screen, automatically completes feature extraction, target recognition and localization, and finally outputs the detection result after post-processing operations such as confidence filtering, non-maximum suppression and multi-scale fusion correction.

Images