Improved face detection method of distance gating image of yolov11s

Abstract

The application discloses a face detection method for a distance-gated image based on an improved YOLOv11s, and adopts a face detection model modified from a YOLOv11s network model to process an input distance-gated image and detect a face. The modification of the YOLOv11s network model comprises the following steps: 1) in the Backbone, for each image dimension reduction module, a spatial-depth conversion CPD-Conv module is used to replace an original convolution module; for a feature extraction module, an attention mechanism module C2DSM is used to replace an original C2PSA module; in the feature extraction of each dimension image in the image dimension reduction process, a feature enhancement and fusion module FEM is used to replace an original feature extraction module C3k2 for the feature extraction of a middle dimension image; 2) in each scale feature map fusion branch of the Neck, a MulitSEAM module is added, respectively, and the outputs of each fusion branch are processed through a MulitSEAM module, and then subsequent processing is performed; and 3) in the detection head Detect of the Head, a repulsion loss is added in cooperation with the MulitSEAM module.

Description

Technical Field

[0001] This invention relates to computer vision technology, specifically an improved YOLOv11s distance-gated image face detection method. Background Technology

[0002] Since its origins in the 1960s, laser range-gated imaging technology has received widespread attention both domestically and internationally with the development of science and technology. Compared to conventional imaging equipment, laser range-gated imaging systems offer advantages such as longer detection range, nighttime imaging capability, and the ability to penetrate windows and smoke to a certain extent. They can also operate normally in adverse weather conditions such as rain, snow, and fog. Combining laser range-gated imaging technology with target detection technology enables high-precision target detection and identification in complex environments. This makes laser range-gated imaging a promising candidate for applications in underwater exploration, fire rescue, hostage search and rescue, and security reconnaissance. .

[0003] Object detection is an important task in computer vision, aiming to automatically locate and identify target objects from images. With technological advancements, object detection methods have evolved from traditional manual feature extraction to deep learning models. - [ The earliest methods were object detection methods based on handcrafted features, which held an important position in traditional computer vision. These methods primarily relied on manually designed feature descriptors to represent local information about objects. A representative method is HOG. SVM Adaboost These types of detection methods are suitable for handling simple targets. Their drawbacks are that, due to their reliance on manually designed features, they are computationally intensive and have poor robustness.

[0004] With the development of deep learning, object detection methods based on deep learning have gradually become mainstream. These methods are further divided into two main categories: two-stage detection and single-stage detection. A representative algorithm for two-stage object detection is R-CNN. SPP-Net Faster R-CNN FPN However, it has drawbacks such as slow inference speed, high model complexity, and dependence on the quality of candidate regions.

[0005] Representative single-stage object detection algorithms include: YOLO SSD , Transformer-based Methods such as DETR In 2015, Joseph Redmon et al. proposed YOLO. The concept of object detection, for the first time, transformed the object detection problem into a regression problem, completing the detection of all objects in an image through a single forward propagation. SSD (Simulated Segmentation Detection), proposed by Wei Liu et al. in 2016, aims to improve inference speed while maintaining high accuracy. It avoids the complexity of multi-stage processing in traditional methods by performing detection directly on multi-layer feature maps. In 2020, Glenn Jocher's Ultralytics team released YOLOv5. Its modular PyTorch implementation, along with its advantages of dynamic anchor optimization and Mosaic data augmentation, has become the "industry standard" in object detection, driving the rapid deployment of AI technology in real-world scenarios. (Ni Guangxing) By fusing the improved YOLOv5 with the Mediapipe method and reconstructing the C3 module using FastNet, the training time cost was reduced and the robustness of recognition was increased. (Song Su) The backbone network of YOLOv5 was optimized and reorganized, and the originally coupled detection heads were decoupled, resulting in a lightweight decoupling head. This improved detection accuracy and convergence speed, while also allowing for better control of the number of model parameters. Then, in 2023, the Ultralytics team developed YOLOv8. The introduction of modules such as C2f effectively combines high-level features with contextual information, significantly improving detection accuracy. (Tian Qing) Wang Ying et al. used a large convolutional kernel C2f-DSF to improve the ability to capture global information, and employed the upsampling operator Dysample to significantly reduce feature information loss. (Min Feng) They incorporated the concept of deep supervision into the detection head and proposed a shallow hybrid pooling downsampling module and a deep max pooling downsampling module. Experiments on the public dataset DUO showed a 2.5 percentage point improvement in mAP@50 compared to the baseline model YOLOv8n. In 2024, YOLOv11 was released again by the Ultralytics team. They designed modules such as C3k2 and C2PSA to enhance the spatial representation capability of features, making significant improvements in both accuracy and efficiency. (Li Bin) For the task of detecting small targets on UAVs, researchers improved the YOLOv11n model by designing the Dilated Feature Pyramid Convolution (DFPC) module and proposing a new feature pyramid structure, which significantly improved the ability to extract global features and fuse multi-scale features.

[0006] In target detection tasks, the clearer the detection image, the better the detection effect. However, due to issues such as imaging principles, laser range-gated images (hereinafter referred to as "range-gated images") have certain defects, such as lack of color information, low resolution, few texture features, and certain noise, which make the detection task more difficult. Summary of the Invention

[0007] As mentioned above, laser range-gated imaging technology has advantages such as long operating distance, the ability to penetrate some objects for imaging, and the ability to operate in adverse weather conditions such as rain, snow, and fog. Addressing the challenges of local feature loss and noise interference caused by complex environments such as windows, obstructions, smoke, and fire, this invention proposes a new face detection algorithm based on the YOLOv11s network and considering the characteristics of range-gated images. This algorithm, through various improvements, primarily enhances the detection performance through vehicle windows, smoke, and fire, thus broadening the application of target detection in range-gated images. The main features are:

[0008] First, a C2DSM module is proposed to replace C2PSA, which can better extract and retain fine-grained features while ensuring training efficiency.

[0009] Secondly, for the neck network, an improved separation and enhanced attention module, MultiSEAM, is adopted to effectively handle the situation where a small part of the region is occluded, thereby improving the feature understanding ability in complex scenes.

[0010] Meanwhile, a CPD-Conv module was added to the backbone network to enhance the feature extraction capability for low-resolution targets; finally, an improved FEM module was used to expand the network's local perception capability and the expression of semantic information of small targets.

[0011] Experiments have verified that the improved YOLOv11 improves the mAP@0.5 and mAP@0.5:0.95 metrics by 3.3% and 1.7% respectively on the gated image dataset, validating the effectiveness of the improvement.

[0012] The specific technical solution of the present invention is as follows:

[0013] An improved YOLOv11s distance gated image face detection method is characterized by improving and modifying the YOLOv11s network model to obtain a face detection model, and using the face detection model to process the input distance gated image to detect faces.

[0014] The YOLOv11s network model consists of: a backbone network, which extracts features from the input image; a neck network, which connects the backbone network and the head network's detection head, used to fuse feature maps of different scales; and a head network, which is used to perform target localization and classification tasks based on the feature maps extracted by the backbone network and the neck network.

[0015] The modifications to the YOLOv11s network model include:

[0016] 1) In Backbone:

[0017] For each image dimensionality reduction module, the original convolution module is replaced with the spatial-depth transform CPD-Conv module;

[0018] For the feature extraction module, the original C2PSA module is replaced with the attention mechanism module C2DSM.

[0019] During the image dimensionality reduction process, when extracting features from images of various dimensions, the feature enhancement and fusion module FEM is used to replace the original feature extraction module C3k2 for the feature extraction of mid-dimensional images.

[0020] 2) In the fusion branches of the feature maps at each scale of Neck, add a MulitSEAM module. The output of each fusion branch is processed by a MulitSEAM module before further processing.

[0021] 3) In the Head's Detect module, a rejection loss was added in conjunction with the MultitSEAM module.

[0022] The Backbone structure consists of a ConV convolution module, the first to fourth image dimensionality reduction modules, an SPPF pooling module, and a C2DSM module, which are connected in series.

[0023] The CPD-Conv module of the first, second, and fourth image dimensionality reduction modules is followed by the C3k2 module; the CPD-Conv module of the third image dimensionality reduction module is followed by the FEM module.

[0024] The input image is processed by Backbone to obtain features of the image in different dimensions.

[0025] In Neck, the processing procedures for the small, medium, and large-scale feature map fusion branches are as follows:

[0026] 1) The low-dimensional image features obtained by the C2DSM module are first concatenated with the features of the medium-scale feature map in the small-scale feature map fusion branch, then processed by the C3k2 module, and finally processed by the MulitSEAM module to obtain the small-scale feature map.

[0027] 2) After upsampling, the low-dimensional image features are first concatenated with the mid-dimensional image features obtained by the FEM module in the mid-scale feature map fusion branch, and then processed by the C3k2 module; the processing result is concatenated with the features of the large-scale feature map, and then processed by the C3k2 module, and finally by the MulitSEAM module to obtain the mid-scale feature map.

[0028] 3) After upsampling, the processing result described in 2) is first concatenated with the features of the high-dimensional image obtained by the second image dimensionality reduction module in the large-scale feature map fusion branch, then processed by the C3k2 module, and finally obtained by the MulitSEAM module to obtain the large-scale feature map.

[0029] The main innovations of this invention are as follows:

[0030] a) In order to effectively recover some of the information lost by the traditional attention mechanism, a C2DSM module was designed to replace the original C2PSA module, so as to better distinguish between strong and weak attention signals;

[0031] b) An improved MulitSEAM module was added to enhance the feature response of the unoccluded area to compensate for the response loss of the occluded area, thereby improving the recognition accuracy of targets in complex scenes where they are partially occluded.

[0032] c) To address the low resolution of distance-gated images, a Channel-Preserved Downsampling (CPD) module is proposed. This module replaces some of the original network's convolutional modules, preserving the discriminative feature information while downsampling the feature mapping.

[0033] d) To enrich local texture information, expand the network's local perception capability, the FEM feature enhancement module was improved, and the semantic information expression of small targets was enhanced. Attached Figure Description

[0034] Figure 1 This is a diagram of the improved YOLOv11s algorithm structure;

[0035] Figure 2 This is a schematic diagram of the CPD-Conv module when scale = 2;

[0036] Figure 3 This is a schematic diagram of the DSM, in which:

[0037] Vs and Vo: Value Vectors, corresponding to the processing results of same-sign and different-sign interactions, respectively.

[0038] Matmul: refers to matrix multiplication.

[0039] ⊙: Element-wise multiplication, which means multiplying corresponding elements of two vectors.

[0040] Figure 4 It is the overall framework of FEM;

[0041] Figure 5 It is the network structure of MultiSeam;

[0042] Figure 6 It is the network structure of CSMM;

[0043] Figure 7 This is a demonstration of the testing results;

[0044] Figure 8 This is a comparison of test results for gating images in complex scenes;

[0045] Figure 9 This is a diagram of the C2DSM module structure. Detailed Implementation

[0046] The present invention will be further described below with reference to specific embodiments.

[0047] 1. Improved YOLOv11s object detection algorithm

[0048] The improved YOLOv11s distance-gated image target detection algorithm of this invention mainly consists of the following three parts:

[0049] (1) Backbone: The core part of the model, whose main function is to extract features from the input image. First, the CPD-Conv module is used to replace some of the original convolutional modules to deal with low resolution. Second, the FEM module is inserted in the middle of the network to expand the network's local perception ability and enhance the expression of semantic information of small objects. In addition, the original C2PSA module is improved and upgraded to C2DSM to better distinguish strong and weak attention signals and extract and retain fine-grained features.

[0050] (2) Neck network: This is the key part connecting the backbone network and the detection head. Its main function is to fuse feature maps of different scales to enhance the model's ability to detect targets at multiple scales. A three-layer MultiSEAM module is added to enhance the feature response of unoccluded areas to cope with occlusion.

[0051] (3) Head network: Its main function is to use the feature maps extracted from the backbone network and neck network, and add repulsion loss (rel) in conjunction with the MultiSEAM module to complete the target localization and classification tasks.

[0052] The improved YOLOv11s network structure is as follows: Figure 1 As shown.

[0053] 1.1 Spatial-Depth Transformation Module CPD-Conv

[0054] While Convolutional Neural Networks (CNNs) have achieved great success in many computer vision tasks, such as image classification and object detection, their performance degrades rapidly when processing low-resolution images or small objects. This is due to a common design flaw in existing CNN architectures: the use of strided convolutions or pooling layers, which leads to the loss of fine-grained information and poor feature representation learning. To address this issue, this invention introduces a novel CNN building block, CPD-Conv. CPD-Conv replaces the original strided convolutional and pooling layers. It consists of a Space-to-Depth (CPD) layer and a non-strided convolutional layer. The CPD layer reduces the spatial dimension of the feature map through specific slicing and concatenation operations, while preserving information in the channel dimension, thus avoiding information loss. The non-strided convolutional layer reduces the number of channels added by the CPD operation and utilizes learnable parameters for feature extraction, solving the problem of fine-grained information loss in low-resolution images and small object detection tasks. The CPD-Conv module mainly consists of two layers: a Space-to-Depth layer and a non-strided convolutional layer. Its structural diagram is shown below. Figure 2 As shown.

[0055] The core idea of ​​the CPD layer is to reduce the spatial dimensionality of the feature map while retaining information in the channel dimension, thereby avoiding information loss. When the scale is 2, the specific operation is as follows:

[0056] Suppose we input a feature map with a size of , where L is the spatial dimension (height and width). This refers to the number of channels. Through a "space-to-depth" operation, the input feature map is decomposed into four sub-feature maps, each with a size of [size missing]. However, they all extract information from different locations in the original feature map. Specifically:

[0057] (1)

[0058] (2)

[0059] (3)

[0060] (4)

[0061] Next, the four sub-feature maps are concatenated along the channel dimension to form a new feature map with a size of [size missing]. The spatial dimension is halved, while the channel dimension is increased to four times its original size. This step downsamples the feature map by transferring spatial information to the channel dimension, and all information is retained in the channel dimension, thus preventing any loss of information. Finally, the concatenated feature map is further processed by a convolutional layer with a stride of 1, and its output size is [size missing]. ,in This refers to the number of output channels of the convolutional layer, generally speaking. This step reduces the number of channels through convolution operations, while extracting features using learnable parameters and preserving important feature information.

[0062] 1.2 Improved Attention Mechanism Module C2DSM

[0063] Traditional softmax-based attention mechanisms face high computational complexity when processing long sequence data. The problem of computational complexity (CLC) limits its practicality in resource-constrained environments. To address this issue, numerous researchers have developed linear attention mechanisms, which reduce computational complexity to linear levels by kernelizing feature maps. However, linear attention is still less expressive than softmax-based attention because it loses information during the approximation process, resulting in reduced discriminative power of the attention map, especially in gated images with complex scenes.

[0064] To address these issues, this invention designs a C2DSM module, the main component of which is DualSignFormer. This is a novel Vision Transformer that improves the performance and efficiency of models in handling visual tasks by introducing a polarity-aware linear attention mechanism. This invention combines it with the C2PSA module in YOLOv11 to derive the improved C2DSM module, which enhances feature discrimination and the model's expressive power, particularly in the detection performance of small targets and complex backgrounds, while also extracting and preserving fine-grained features.

[0065] The network structure of the C2DSM module is as follows: Figure 9 As shown. The structure of DSM (DualSignFormer) is as follows: Figure 3 As shown. Because polarity-aware attention decomposes the query and key vectors into their positive and negative components, it is translated as dual-sign.

[0066] The C2DSM module is based on Dual Sign Former and replaces the self-attention mechanism in the C2PSA module of YOLOv11s with DSM. In other words, C2DSM is an improvement on C2PSA in YOLOv11s, modifying and replacing the PCA module in C2PSA with the DSM module.

[0067] When handling negative components, DualSignFormer employs a polarity-aware linear attention mechanism to decompose the query and key vectors into their positive and negative components. This decomposition mechanism considers the influence of positive and negative similarity on the attention weights separately. Specifically, for the query vector q and key vector k, they can be decomposed as follows:

[0068] (5)

[0069] in, and Let represent the positive and negative components of q, respectively, and the same applies to k. Substituting these decompositions into the inner product of q and k, we get:

[0070] (6)

[0071] The first two terms capture the similarity between components with the same sign, while the latter two terms represent the interaction between components with opposite signs. Then, the polarity-aware attention mechanism separates them according to the polarity of q and k and independently calculates their interaction. The attention weights are calculated as shown in Equation (7), which recovers the information embedded in the positive and negative parts.

[0072] (7)

[0073] Finally, the output vector is shown in formula (8).

[0074] (8)

[0075] This formula handles the interaction between these two parts separately, thereby capturing richer relationship information. Specifically, it is divided into two types: same-sign interaction and opposite-sign interaction. Same-sign interaction refers to... and The dot product calculation with the same sign interaction, the result is obtained through... Scaling. The same applies to interactions with different signs. In this way, DualSignFormer can improve the model's expressive power and the discriminative power of the attention map, capturing more details and differences. Therefore, this mechanism is particularly helpful in handling data with complex relationships and is suitable for object detection problems in complex scenarios.

[0076] 1.3 Feature Enhancement and Fusion Module (FEM)

[0077] In complex background environments, especially in gated images, targets are easily confused with the background, making detection difficult. Traditional models may fail to fully capture the subtle features of small targets, perform poorly when fusing feature maps of different scales, and deep networks are prone to losing details, resulting in insufficient feature representation capabilities, difficulty in distinguishing targets from background, and limited detection performance. Accuracy is further affected when faced with image degradation (such as blur, noise, fog, etc.). Traditional methods such as RFBNet and MobileNet suffer from drawbacks such as weak feature extraction capabilities and structural redundancy.

[0078] To address this issue, this invention improves the Feature Enhancement Module (FEM). It is inserted into the backbone network and connected to the network neck.

[0079] The aim is to obtain rich local contextual information through multi-branch dilated convolutional structures, thereby enhancing the network's ability to perceive targets. The network structure of the FEM module... Figure 4 As shown.

[0080] Its mathematical expression is as follows:

[0081] (9)

[0082] in, For branch output, This is a series splicing operation. This is an element-wise addition operation of the feature maps. It is the input feature map. This is the output of the FEM module.

[0083] The main features and functions of the FEM module include the following:

[0084] (1) Multi-branch parallel structure: FEM adopts a multi-branch structure that includes standard convolution and atrous convolution. This operation can capture features at different scales and a wider range of contextual information, thus enhancing perception capabilities.

[0085] (2) Feature fusion: The output of FEM is the fusion result of the output feature maps of multiple branches. This method enables the detection network to effectively enhance the local perception capability of the input feature map, providing richer feature representations for subsequent feature fusion and global context modeling.

[0086] 1.4 Improved Neck Module MultiSEAM

[0087] While deep convolutional networks have significantly improved face detection, in gated scenarios, faces may be obscured by other objects (such as masks, glasses, hair) or other facial features (such as mutual occlusion in dense crowds), making it difficult for traditional detectors to accurately locate and recognize faces. Occlusion causes partial loss of facial features, affecting the recall and precision of the detector, especially in complex scenes. Early occlusion handling methods such as CascadeCNN and Context Modeling have significant drawbacks. To address this issue, this invention improves upon SEAM. The Separated and Enhancement Attention Module uses a multi-branch structure and introduces a repulsion loss. (Repulsion Loss, rel) designed the MultiSeam module.

[0088] The core function of the MultiSeam module is to address feature loss and localization bias caused by occlusion by dynamically enhancing the feature response of effective regions through an attention mechanism. Its main component is the Channel and Spatial Mixing Module (CSMM), which aims to enhance the model's ability to detect occluded faces through depthwise separable convolutions and channel / spatial mixing mechanisms. This module uses methods such as depthwise convolution (Dconv), pointwise convolution (Pconv), dilated convolution, residual connections, and channel / spatial information fusion to achieve multi-scale feature fusion, strengthening the features of unoccluded regions and compensating for the response loss of occluded parts, thereby improving the feature response in occluded scenes. Its network structure diagram is shown below. Figure 5 As shown.

[0089] For the input features, MultiSEAM uses multiple CSMM branches with different patch sizes (e.g., 5×5, 7×7, 9×9) to capture local features under different receptive fields. Each CSMM branch contains depthwise convolution (DConv), pointwise convolution (Pconv), and residual connection structures, enhancing feature representation while reducing the number of parameters. Then, average pooling is used to fuse the features from the three branches and the original input, preserving multi-scale information. Finally, fully connected layers generate channel attention weights using the sigmoid function and exponential scaling to expand the dynamic range of the weights, enhancing the response to key information.

[0090] Figure 6 This demonstrates the specific structure of the CSMM module. In this module, the idea of ​​separable convolution is employed, separating depthwise convolution and pointwise convolution into two branches for feature extraction, which are then fused together. Dilated convolution is then used to further expand the receptive field without increasing the number of parameters, achieving joint channel and spatial optimization.

[0091] During detection, traditional non-maximum suppression (NMS) may lead to missed detections due to occlusion. To address the issue of occlusion of targets in gated images, this invention introduces a repulsion loss function. This loss function constrains the distribution of predicted bounding boxes, keeping them away from other ground truth boxes (RepGT) and other predicted bounding boxes (RepBox).

[0092] The repulsion loss is divided into two parts, each addressing different situations: RepGT Loss: makes the current predicted bounding box... It is far from other ground truth (GT) boxes. Its mathematical formula is as follows.

[0093] (10)

[0094] Smoothing function The mathematical formula is shown in (1.11), which applies a stronger penalty to high overlap values.

[0095] (11)

[0096] The meaning it represents is a prediction frame. With the nearest non-matching ground truth box The overlap ratio.

[0097] RepBox Loss: This function moves different predicted bounding boxes in the same image further apart, preventing accidental deletion by NMS. The formula is shown below.

[0098] (12)

[0099] Only predictive bounding box pairs with IoU > 0 are computed. To prevent the denominator from being zero, the repulsion loss compensates for the shortcomings of the pure attention mechanism through geometric constraints. Together with MultiSEAM's enhancement of occluded region features, it synergistically improves face detection performance. The total loss function of YOLOv11 can be expressed as:

[0100]

[0101] 2. Experimental Results and Analysis

[0102] 2.1 Dataset

[0103] Due to the lack of gated image data for complex transmission scenarios (such as through windows, dense smoke, and flames), this experiment constructed its own gated image dataset. The data was acquired using a laser distance-gated camera and contains 3,180 valid images, with 90% showing through vehicle windows and 10% showing through dense smoke and flames. All images include human faces and are labeled using the LabelImg tool (label file format: .xml). The dataset was randomly divided into a training set (2,226 images), a validation set (636 images), and a test set (318 images) in a 7:2:1 ratio, ensuring a uniform distribution of samples from different scenarios across the subsets.

[0104] 2.2 Experimental Environment and Parameter Settings

[0105] This experiment used a Windows 11 system with an AMD Ryzen 7 6800H CPU with Radeon Graphics, an Nvidia GeForce RTX 4090 graphics card with 8GB of VRAM, Python version 3.9, PyTorch version 2.5.1+cu118, and CUDA version 12.0. The batch size was set to 16, the training image size to 640×640, the epochs to 300, no pre-trained weights used, the initial learning rate was 0.01, and SGD was selected as the optimizer.

[0106] 2.3 Model Evaluation Indicators

[0107] The experiment used mAP@0.5 and mAP@0.5:0.95 as the main evaluation indicators to reflect the overall performance of the model under relaxed and strict IoU thresholds, respectively.

[0108] mAP@0.5 is the mean Average Precision calculated with an Intersection over Union (IoU) threshold of 0.5. It represents the average detection accuracy of the model for all classes when the IoU between the predicted bounding box and the ground truth bounding box is ≥ 0.5.

[0109] mAP@0.5:0.95 is the average of mAP calculated at multiple IoU thresholds (from 0.5 to 0.95, with a step size of 0.05), which better reflects the model's sensitivity to the bounding box position. It requires the model to perform well under different levels of localization accuracy.

[0110] Taking the COCO standard as an example, the calculation process is divided into single-class AP calculation, multi-IoU thresholding, and multi-class averaging. Specifically, the predicted bounding boxes are first sorted by confidence level, and the precision under different recall levels is calculated.

[0111] (13)

[0112] (14)

[0113] Plot the PR curve and calculate the area under the curve (AP).

[0114] (15)

[0115] The representative recall rate is The accuracy value at that time was calculated, and then the above process was repeated for each IoU threshold (0.5~0.95) to obtain 10 AP values. The average value was then taken to obtain mAP@0.5:0.95.

[0116] (16)

[0117] It is the first AP values ​​for each category This represents the total number of categories.

[0118] 2.4 Analysis and Comparison of Experimental Results

[0119] like Figure 7 As shown in the figure, the detection performance of the improved YOLOv11s is demonstrated, with all test images being non-training images. To verify the improvement, this experiment also used other networks for comparative testing on a self-built gated test dataset, including YOLO series networks, SSD, Retinanet, EfficientDet, and Swin-Tran-sformer. Experimental results show that the proposed method achieves mAP@0.5 and mAP@0.5:0.95 of 93.6% and 66.4%, respectively, representing improvements of 3.3% and 1.7% compared to the original network. Specific comparative experimental results are shown in Table 1.

[0120] Table 1 Comparison of detection performance on gated datasets

[0121]

[0122] 2.5 Ablation Experiment

[0123] This experiment conducted systematic ablation experiments on a self-built gated dataset to quantitatively evaluate the contribution of each innovative module to model performance. Using YOLOv11s as the base model, the experiment employed a progressive strategy, sequentially using the following improved modules: first, replacing the original conv module with the CPD-Conv module; second, upgrading the C2PSA layer in the backbone network to the C2DSM module; next, adding an FEM structure in the middle of the network; and finally, introducing the dynamic detection head MultiSEAM. Table 2 details the impact of each improvement stage on detection accuracy (mAP), parameter count (Param / M), and computational efficiency (GFLOPs), where the baseline represents the performance of the original YOLOv11s model. This invention verifies the progressive improvement effect of each module on model performance through systematic ablation experiments. Based on YOLOv11s (9.4M / 21.3GFLOPs / mAP@0.5:0.95=64.7%), by gradually adding CPD-Conv, C2DSM, FEM and MultiSEAM modules, the model achieved improvements of 3.3% (to 93.6%) in mAP@0.5 and 1.7% (to 66.4%) in mAP@0.5:0.95 with only a 13.6% increase in computational cost.

[0124] Specifically, CPD-Conv delivers a 0.6% initial accuracy improvement with minimal computational cost (+0.1 GFLOPs); C2DSM further enhances multi-scale feature fusion, improving mAP@0.5 by 0.8%; the FEM module significantly enhances robustness in complex scenes (mAP@0.5: 0.95 + 0.5%); and finally, MulitSEAM contributes the largest gain through its multi-scale attention mechanism (single module improvement of 40%). All improvements achieve Pareto optimality in accuracy and efficiency while maintaining a lightweight parameter set (total parameter count +9.6%), with significant improvements in detection of window-through, smoke-through, fire-through, and occluded scenes.

[0125] Table 2 Ablation Experiment

[0126]

[0127] 2.6 Visualization of Experimental Results

[0128] like Figure 8As shown, this paper compares the performance of the original YOLOv11s network (middle column) and the improved YOLOv11s network (right column) on a self-built gated dataset. The green boxes indicate the differences. The four images shown include complex scenes such as windowed images, occlusions, and dense crowds. The left column contains the original images. Experimental results show that the original YOLOv11s network suffers from missed detections and duplicate detection boxes when dealing with gated images in complex scenes, while the improved network can accurately detect faces. This indicates that the original YOLOv11s network cannot accurately and effectively complete face detection tasks in complex scenes with occlusions. The improvements made to the four modules CPD-Conv, C2DSM, FEM, and MultiSEAM in this experiment enable the improved network to effectively handle these complex scenes and accurately complete face detection tasks.

[0129] 3. Conclusion

[0130] To address the challenge of face detection in complex scenarios such as windowed, occluded, smoke-filled, and fire-filled gated images, this invention first addresses the low resolution of gated images by replacing the original conv layers with CPD-Conv. The spatial-to-depth pass-through reduces the spatial dimensionality of the feature map while preventing information loss, and the non-stretch convolutional layers reduce the number of channels added by CPD operations, thus resolving the loss of fine-grained information in low-resolution images and small object detection tasks. Secondly, to improve upon the shortcomings of traditional attention mechanisms, a polarity-aware linear attention mechanism is introduced, and the C2PSA module is modified to design the C2DSM module. This improves the model's expressive power and the discriminative power of the attention map, capturing more details and differences, which is helpful for handling data with complex relationships. Then, an improved FEM module is used, enabling the detection network to effectively enhance the local perception capability of the input feature map, providing richer feature representations for subsequent feature fusion and global context modeling. Finally, the MultiSEAM module enhances the features of unoccluded regions, compensating for the response loss of occluded parts. With the development and popularization of artificial intelligence technology and laser range-gated imaging technology, the improved YOLOv11s network has broad application prospects in underwater exploration, fire rescue, hostage search and rescue, security reconnaissance and other fields.

[0131] References:

[0132] [1] Xu Xiaowen, Guo Jin, Yu Qianyang, et al. Key technologies for laser distance-gated imaging [J]. Laser Technology, 2003, (06): 603-605.

[0133] XU XW, GUO J, YU QY, et al. Key technique of l-aser range gatedimaging[J]. Laser Technology, 2003, (06): 603-605.

[0134] [2] Zhu Xiaopeng, Liu Jiqiao, He Yan, et al. 532nm laser distance-gated imaging system [J]. Infrared and Laser Engineering, 2012, 41(02): 358-362.

[0135] ZU XP, LIU JQ, HE Y, et al. Range gated imaging lidar at wavelength of 532nm[J]. Infrared and Laser Engineering, 2012, 41 (02): 358-362.

[0136] [3] Wang Shouzeng, Sun Feng, Zhang Xin. Research progress on laser illumination distance gating imaging technology [J]. Infrared and Laser Engineering, 2008, 37(S3): 95-99.

[0137] WANG SZ, SUN F, ZHANG X. Development of laser illuminating range-gated imaging technique[J].Infrared and Laser Engineering, 2008, 37 (S3): 95-99.

[0138] [4]Z. Zou, K. Chen, Z. Shi, Y. Guo and J. Ye. Object Detection in 20Years: A Survey[C]. in Proceedings of the IEEE, vol. 111, no. 3, pp. 257-276, March 2023.

[0139] [5]U. DWIVEDI, K. JOSHI, SK SHUKLA and AS RAJAWAT. An Overview of Moving Object Detection Using YOLODeep Learning Models[C]. 2024 2ndInternational Conference on Disruptive Technologies (ICDT), Greater Noida,India, 2024, pp. 1014-1020.

[0140] [6] Li Qiong, Kao Yueying, Zhang Ying, et al. A review of research on target detection for UAV aerial images [J]. Acta Graphica Sinica, 2024, 45(06): 1145-1164.

[0141] LI Q, KAO YY, ZHANG Y, et al. Review on object-ct detection in UAVaerial images[J]. Journal of Graphics, 2024, 45 (06): 1145-1164.

[0142] [7] Zhu Ziwen, Song Xiaoou, Cui Wei, et al. A review of target detection using visible light-infrared image fusion [J / OL]. Computer Engineering and Applications, 1-26 [2025-04-01].

[0143] ZHU ZW, SONG XO, CUI W, et al. Visible and Infrared Image Fusion for Intelligent Object Detection: A Review[J / OL]. Computer Engineering andApplications, 1-26[2025-04-01].

[0144] [8] Wang Ning, Zhi Min. A review of research on single-stage general object detection algorithms under deep learning [J / OL]. Computer Science and Exploration, 1-32 [2025-04-01].

[0145] WANG N, ZHI M. A Review of One-Stage Universal Object DetectionAlgorithms in Deep Learning[J / OL]. J-ournal of Frontiers of Computer Scienceand Technology,1-32[2025-04-01].

[0146] [9]C. Yang, Z. Gao, Y. Hao and R. Yang. Survey of Lightweight ObjectDetection Model Methods[C], 2024 4th International Conference on ArtificialIntelligence, Robotics, and Communication (ICAIRC), Xiamen, China, 2024, pp.101-113.

[0147]

[10] N. Dalal and B. Triggs. Histograms of oriented gradients forhuman detection[C]. 2005 IEEE Computer Society Conference on Computer Visionand Pattern Recognition (CVPR'05), San Diego, CA, USA, 2005, pp. 886-893 vol.1

[0148]

[11] Cortes, C., Vapnik, V. Support-vector networks[J]. Mach Learn 20,273–297 (1995).

[0149]

[12] Yoav Freund, Robert E Schapire. A Decision T-heoreticGeneralization of On-Line Learning and an Ap-plication to Boosting. Journalof Computer and System Sciences, 1997, Volume 55, Issue 1, ISSN 0022-0000.

[0150]

[13] R. Girshick, J. Donahue, T. Darrell and J. Malik. Rich FeatureHierarchies for Accurate Object Detection and Semantic Segmentation[C]. 2014IEEE Conference on Computer Vision and Pattern Recognition, Columbus, OH,USA, 2014, pp. 580-587.

[0151]

[14] K. He, X. Zhang, S. Ren and J. Sun. Spatial Py-ramid Pooling inDeep Convolutional Networks for Vis-ual Recognition[J]. IEEE Transactions onPattern Analy-sis and Machine Intelligence, 1 Sept 2015, vol. 37, no. 9, pp.1904-1916.

[0152]

[15] REN S Q,HE K M,GIRSHICK R,et al.Faster RCNN : towards real- timeobject detection with region proposalnetworks[C] / / Advances in NeuralInformation Processing Systems,2015:91-99.

[0153]

[16] S. Ren, K. He, R. Girshick and J. Sun, Faster R-CNN: TowardsReal-Time Object Detection with Region Proposal Networks[J], in IEEETransactions on Pattern Analysis and Machine Intelligence, vol. 39, no. 6,pp. 1137-1149, 1 June 2017.

[0154]

[17] J. -H. Xu, J. -P. Li, Z. -R. Zhou, Q. Lv and J. Luo. A Survey ofthe Yolo Series of Object Detection Algorithms[C]. 2024 21st InternationalComputer Conference on Wavelet Active Media Technology and InformationProcessing (ICCWAMTIP), Chengdu, China, 2024, pp. 1-6.

[0155]

[18] LIU W,ANGUELOV D,ERHAN D,et al. SSD:single shot multibox detector[C] / / EuropeanConference on Computer Vision,2016:21-37.

[0156]

[19] Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A.,Zagoruyko, S. (2020). End-to-End Object Detection with Transformers[C]. In:Vedaldi, A., Bischof, H., Brox, T., Frahm, JM. (eds) Computer Vision – ECCV2020. ECCV 2020. Lecture Notes in Computer Science(), vol 12346. Springer,Cham.

[0157]

[20] J. Redmon, S. Divvala, R. Girshick and A. Farhadi. You Only LookOnce: Unified, Real-Time Object Detection[C]. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Las Vegas, NV, USA, 2016, pp.779-788, doi: 10.1109 / CVPR.2016.91.

[0158]

[21] Ni Guangxing, Xu Hua, Wang Chao. Research on gesture recognition by integrating improved YOLOv5 and Mediapipe [J]. Computer Engineering and Applications, 2024, 60(07): 108-118.

[0159] NI GX, XU H, WANG C. Research on Gesture Recognition Based on Improved YOLOv5 and Mediapipe[J]. Computer Engineering and Applications, 2024, 60 (07): 108-118.

[0160]

[22] Song Su, Wang Fangzheng, Gao Jianan, et al. Target detection model for UAV aerial photography based on multi-path dynamic convolution YOLOv5[J]. Modern Electronics Technology, 2025, 48(07): 72-78. DOI:10.16652 / j.issn.1004-373x.2025.07.011.

[0161] SONG S, WANG FZ, GAO JA, et al. YOLOv5 UAVaerial photography objectdetection model based on mul-tipath dynamic convolution[J]. ModernElectronics Tech-nique,2025,48(07):72-78.DOI:10.16652 / j.issn.1004-373x.2025.07.011.

[0162]

[23] VARGHESE, REJIN, SAMBATH M. YOLOv8: A Novel Object DetectionAlgorithm with Enhanced Performance and Robustness. 2024 International Conference on Advances in Data Engineering and Intelligent Computing Systems (ADICS) (2024): 1-6.

[0163]

[24] Tian Qing, Wang Ying, Zhang Zheng, et al. Improved YOLOv8n gated image target detection algorithm [J]. Computer Engineering and Applications, 2025, 61(02): 124-134.

[0164] TIAN Q, WANG Y, ZHANG Z, et al. Improved YOLOv8n for Gated ImagingObject Detection Algorithm[J]. Computer Engineering and Applications, 2025,61(02):124-134.

[0165]

[25] Min Feng, Zhang Yuwei, Liu Yuhui, et al. Improved lightweight underwater biological detection model for YOLOv8 [J]. Computer Engineering and Applications, 2025, 61(06): 96-105.

[0166] MIN F, ZHANG YW, LIU YH, et al. Improving Lig-htweight Underwater Biological Detection Model of YO-LOv8[J]. Computer Engineering andApplications, 2025,61(06):96-105.

[0167]

[26] Li Bin, Li Shenglin. Improved YOLOv11n algorithm for small target detection in UAVs [J / OL]. Computer Engineering and Applications, 1-11 [2025-04-02].

[0168] LI B, LI S L. Improved YOLOv11n Small Object Detection Algorithm inUAV View[J / OL]. Computer Engineering and Applications, 1-11[2025-04-02].

[0169]

[27] SUNKARA R, LUO T. No More Strided Convolutions or Pooling: A NewCNN Building Block for Low-Resolution Images and Small Objects[J]. 2022Computer Science(), vol 13715. Springer, Cham.

[0170]

[28] Weikang Meng and Yadan Luo and Xin Li and Dongmei Jiang and ZhengZhang. PolaFormer: Polarity-aware Linear Attention for Vision Transformers[C]. The Thirteenth International Conference on Learning Representations,2025

[0171]

[29] Y. Zhang, M. Ye, G. Zhu, Y. Liu, P. Guo and J. Yan,。FFCA-YOLO forSmall Object Detection in Remote Sensing Images[C]. IEEE Transactions onGeoscience and Remote Sensing, vol. 62, pp. 1-15, 2024, Art no. 5611215.

[0172]

[30] YU, Z P et al. Yolo-facev2: A scale and occlusion a-ware facedetector[J]. Pattern Recognition, Volume 155, 2024,110714.

Claims

1. An improved YOLOv11s distance-gated image face detection method, characterized by: A face detection model was obtained by modifying the YOLOv11s network model, and the face detection model was used to process the input distance gating image to detect faces. The YOLOv11s network model consists of: a backbone network, which extracts features from the input image; a neck network, which connects the backbone network and the head network's detection head, used to fuse feature maps of different scales; and a head network, which is used to perform target localization and classification tasks based on the feature maps extracted by the backbone network and the neck network. The modifications to the YOLOv11s network model include: 1) In Backbone: For each image dimensionality reduction module, the original convolution module is replaced with the spatial-depth transform CPD-Conv module; For the feature extraction module, the original C2PSA module is replaced with the attention mechanism module C2DSM. During the image dimensionality reduction process, when extracting features from images of various dimensions, the feature enhancement and fusion module FEM is used to replace the original feature extraction module C3k2 for the feature extraction of mid-dimensional images. 2) In the fusion branches of the feature maps at each scale of Neck, add a MulitSEAM module. The output of each fusion branch is processed by a MulitSEAM module before further processing. 3) In the Head's Detect module, a rejection loss was added in conjunction with the MultitSEAM module.

2. The improved YOLOv11s distance-gated image face detection method according to claim 1, characterized in that: The Backbone structure consists of a ConV convolution module, the first to fourth image dimensionality reduction modules, an SPPF pooling module, and a C2DSM module, which are connected in series. The CPD-Conv module of the first, second, and fourth image dimensionality reduction modules is followed by the C3k2 module; the CPD-Conv module of the third image dimensionality reduction module is followed by the FEM module. The input image is processed by Backbone to obtain features of the image in different dimensions.

3. The improved YOLOv11s distance-gated image face detection method according to claim 1 is characterized in that the processing procedures for the small, medium, and large scale feature map fusion branches in the Neck are as follows: 1) The low-dimensional image features obtained by the C2DSM module are first concatenated with the features of the medium-scale feature map in the small-scale feature map fusion branch, then processed by the C3k2 module, and finally processed by the MulitSEAM module to obtain the small-scale feature map. 2) After upsampling, the low-dimensional image features are first concatenated with the mid-dimensional image features obtained by the FEM module in the mid-scale feature map fusion branch, and then processed by the C3k2 module; the processing result is concatenated with the features of the large-scale feature map, and then processed by the C3k2 module, and finally by the MulitSEAM module to obtain the mid-scale feature map. 3) After upsampling, the processing result in step 2) is first concatenated with the features of the high-dimensional image obtained by the second image dimensionality reduction module in the large-scale feature map fusion branch, then processed by the C3k2 module, and finally obtained by the MulitSEAM module to obtain the large-scale feature map.

4. The improved YOLOv11s distance-gated image face detection method according to claim 1, characterized in that: The C2DSM module is an attention mechanism module; the C2PSA of YOLOv11s is modified by replacing the PCA module in C2PSA with the DSM module to obtain the C2DSM module; In DSM, In DSM, a polarity-aware linear self-attention mechanism is used to decompose the query vector (q) and key vector (k) into their positive and negative components. This decomposition mechanism considers the influence of positive and negative similarity on the attention weights, respectively. Specifically, for the query vector q and key vector k, they can be decomposed as follows: ， in, and These represent the positive and negative components of q, respectively. and These represent the positive and negative components of k, respectively. Substituting them into the inner product of q and k, we get: ， The first two items capture the similarity between components with the same sign, while the latter two items represent the interaction between components with different signs. Subsequently, the polarity-aware attention mechanism separates q and k according to their polarities and independently calculates their interactions. The attention weights are given by the following formula, which recovers the information embedded in the positive and negative parts: ， Finally, the output vector is as follows: 。 5. The improved YOLOv11s distance-gated image face detection method according to claim 1, characterized in that: The CPD-Conv module replaces the corresponding strided convolutional and pooling layers in YOLOv11s; the CPD-Conv module includes a spatial-to-depth layer and a non-strided convolutional layer; the spatial-to-depth layer reduces the spatial dimension of the feature map while retaining information in the channel dimension.

6. The improved YOLOv11s distance-gated image face detection method according to claim 1, characterized in that: The FEM module is a feature enhancement and fusion module. In the FEM module, the input features are processed through three branches. The results of these branches are concatenated and then added to the elements of the input feature map before being output. The three branches are processed as follows: Serialized 1×1 ordinary convolution, 3×1 ordinary convolution, 1×3 ordinary convolution and 3×3 dilated convolution; Serialized 1×1 ordinary convolution, 1×3 ordinary convolution, 3×1 ordinary convolution and 3×3 dilated convolution; Serialized 1×1 ordinary convolution and 3×3 ordinary convolution.

7. The improved YOLOv11s distance-gated image face detection method according to claim 1, characterized in that: The MultiSEAM module is an improvement on SEAM. For the input features, the MultiSEAM module first uses multiple CSMM branches with different patch sizes to capture local features under different receptive fields. Next, the features of each branch and the original input are fused by average pooling; finally, the channel attention weights are generated by a fully connected layer, using the sigmoid function and introducing exponential normalization. In each CSMM branch, the input feature map first undergoes Patch Embedding, then is fused by depthwise convolution (Dconv) and pointwise convolution (Pconv), and finally undergoes dilated convolution, followed by activation and normalization operations to obtain the output.

8. The improved YOLOv11s distance-gated image face detection method according to claim 1, characterized in that... The Head's Detect function introduces a repulsion loss function, which constrains the distribution of the predicted boxes, making the obtained predicted boxes far away from other ground truth boxes (RepGT) and other predicted boxes (RepBox). The rejection loss is divided into two parts depending on the situation: Part 1: RepGT Loss Make the current prediction box The RepGT Loss, which is far removed from other real bounding boxes, is as follows: ， Among them, the smoothing function As shown in the following formula ， Represents the prediction box With the nearest non-matching ground truth box The overlap ratio is as follows: ， Part Two: RepBox Loss To prevent false deletions by NMS, different predicted bounding boxes in the same image are kept far apart. The RepBox Loss formula is as follows: ， Only predictive box pairs with IoU > 0 are computed. Used to prevent the denominator from being zero; The total loss function of YOLOv11 is as follows: ， The loss function of the face detection model is as follows: 。

Images