基于改进YOLO v8n-pose的多视角羊脸关键点检测方法

Multi-view Key Point Detection Method For Sheep Faces Based on Improved YOLO v8n-pose

摘要:

羊脸关键点检测是羊脸识别、面部表情及头部姿态分析技术的重要基础。针对自然养殖环境下羊脸面部朝向多变、尺度不一等检测难点，本研究提出基于YOLO v8n-pose模型改进的YOLO v8n-DCL-pose模型，该模型在保持轻量化结构的基础上显著提升了对多视角、多尺度羊脸关键点检测精度。采用改进的可变形卷积（Deformable convolution network v4)改进c2f模块，利用动态采样机制增强模型在羊脸形变以及羊脸尺度不一情况下的特征自适应能力;在特征融合层新增1个160像素×160像素的小目标检测层，提升模型对远距离小尺度羊脸关键点的检测能力。采用轻量化非对称解耦检测头（Lightweight asymmetric decoupled head，LADH-Head）将原有的任务耦合检测头拆分为独立的分类、回归和IoU（Intersection over union）分支，提高关键点检测精度。在自建数据集上，改进的YOLO v8n-DCL-pose模型在保持计算效率的同时，较YOLO v8n-pose参数量仅增加6×10^5，AP提升5.2个百分点，mAP50提升5.1个百分点，mAP95提升9.9个百分点，表现出更优的关键点检测性能。YOLO v8n-DCL-pose模型能够有效提取不同视角羊脸关键点信息，可为后续羊脸识别等任务提供关键信息支撑。

Abstract:

Key point detection of sheep faces was regarded as a fundamental step for sheep face recognition, facial expression analysis, and head posture estimation.To address the challenges posed by variable orientations and inconsistent scales of sheep faces in natural breeding environments, an improved YOLO v8n-pose network model, named YOLO v8n-DCL-pose, was proposed.This model significantly improved the detection accuracy of key points on multi-view and multi-scale sheep faces while maintaining a lightweight structure.Specifically, several enhancements were made: firstly, an improved deformable convolution module, DCA, was employed to replace the original c2f module.A dynamic sampling mechanism was utilized to enhance the model's adaptability to sheep face deformations and varying scales.Secondly, a 160 pixel×160 pixel small-object detection layer was added at the feature fusion stage to strengthen the model's ability to detect key points on small-scale, long-distance sheep faces.Finally, a lightweight asymmetric decoupled head (LADH-Head) was adopted, in which the original coupled detection head was decomposed into separate branches for classification, regression, and IoU prediction, thereby improving the precision of key point detection.Experiments on a self-built dataset showed that the improved YOLO v8n-DCL-pose model increased the parameter size by only 6×10^5, while maintaining computational efficiency.The overall AP was improved by 5.2 percentage points, mAP50 by 5.1 percentage points, and mAP95 by 9.9 percentage points compared with that of the original model, demonstrating superior performance in key point detection.These results indicated that the proposed YOLO v8n-DCL-pose model served as an effective key point extractor, which was capable of providing critical feature support for downstream tasks such as sheep face recognition.