关注细粒度特征的复杂场景牛只头部位置轻量化检测模型VSF-YOLO研究

Research on Lightweight Detection Model VSF-YOLO for Cattle Head Position in Complex Scenes on Fine-grained Features

摘要:

当前牲畜检测算法常因复杂场景的干扰而面临采集图像模糊、残缺遮挡等问题,导致检测精度降低。同时检测模型通常参数量大,难以部署至便携式设备。为此提出一种轻量化、应对复杂场景和聚焦面部细粒度特征的VSF-YOLO网络架构应用于牛只面部检测。算法首先提出一种VegsNet主干网络,将其作为YOLO v10n的主干网络以减少模型复杂度,降低运行参数;其次,为捕捉图像细粒度特征,提出一种空间感受野卷积高效地捕捉全图关键信息,并设计一种全文特征锚点注意力机制,利用全文锚点特征信息对细粒度特征进行准确抓取;最后,通过优化损失函数增强卷积对细粒度特征的提取,定位关键区域。为检测算法的有效性和实时性,在特定数据集上进行消融实验和与多种经典算法的对比实验,并部署至手机Android 系统进行测试。结果表明,与传统YOLO v10n相比参数量下降30% ,mAP@0.5提升3. 55个百分点,得益于细粒度特征抓取,在相同模糊、遮挡复杂场景下牛脸的检测置信度平均提高36%,且移植到手机应用程序后运行效率提升,相比原模型运行帧率平均提升30%。

Abstract:

In practical applications, current livestock detection algorithms often suffer from issues such as blurred, incomplete, or occluded captured images due to interference from complex scenes, leading to reduced detection accuracy. Additionally, detection models typically have a large number of parameters, making them difficult to deploy on portable devices. To address these challenges, a lightweight VSF-YOLO network architecture designed for cattle facial detection was proposed, which was capable of handling complex scenes and focusing on fine-grained facial features. The algorithm firstly introduced a VegsNet backbone network to replace the original YOLO v10n backbone, thereby reducing model complexity and operational parameters. Secondly, to capture fine-grained image features, a spatial receptive field convolution was proposed to efficiently extract key information from the entire image. Additionally, a full-feature anchor attention mechanism was designed to accurately capture fine-grained features by leveraging anchor point feature information across the entire image. Finally, the loss function was optimized to enhance the convolution's ability to extract fine-grained features and locate critical regions. To validate the algorithm's effectiveness and real-time performance, ablation studies and comparative experiments with various classical algorithms were conducted on a specific dataset. The model was also deployed on an Android mobile system for testing. The results indicated that compared with the traditional YOLO v10n, the parameter count was decreased by 30%, while the average accuracy was increased by 3.55 percentage points. Benefiting from fine-grained feature extraction, the detection confidence for cattle faces was improved by an average of 36% under the same complex scenarios involving blurring and occlusion. Moreover, after being ported to mobile applications, the operational efficiency was increased, with the average frame rate rising by 30% compared with that of the original model.