基于多模块增强策略的肉羊体尺自动测量方法

Automatic Measurement Method for Sheep Body Dimensions Based on Multi-module Enhancement Strategy

摘要:

针对非接触式体尺测量方法采集肉羊体尺时,因羊只体表纹理平滑、姿态多变,导致特征点定位困难、测量基准不一致、表面细节信息易丢失等问题,提出了一种基于多模块增强策略的肉羊体尺自动测量方法。 分别建立羊只侧视与俯视图像数据集各 5 000 幅。 以 YOLO 11n-Pose 为基础模型,在骨干、颈部及检测头中分别引入集成 PConv 的 Faster_Block、动态可变形卷积 DCNv4、RepViTBlock 及 EVCBlock 模块,构建新的关键点检测模型 PDRE-YOLO,基于 PDRE-YOLO 识别羊只肩拐点、肩高点、胸深点、胸高点、臀高点、坐骨点、蹄尖点、左肩点、右肩点 9 个关键点,根据关键点像素值以及坐标转换算法实现体长、体高、胸深、胸高、十字部高、胸宽 6 项体尺的自动测量。 试验结果表明:在侧视数据集上,模型关键点检测 mAP@ 0. 5:0. 95 达到 99. 50% ,较原始 YOLO 11n-Pose 模型提升 1. 8 个百分点;检测精确率为 93. 60% ,提升 3. 8 个百分点。 在俯视数据集上,其 mAP@ 0. 5:0. 95 同样为 99. 50% , 提升 1. 5 个百分点;检测精确率为 92. 60% ,提升 2. 0 个百分点。 模型检测的关键点经坐标转换后所得体尺数据与人工测量结果对比,平均绝对误差分别为 3. 94、3. 23、2. 04、1. 06、5. 47、1. 42 cm, 平均相对误差分别为 7. 63% 、 5. 87% 、8. 75% 、4. 02% 、10. 22% 、9. 68% 。

Abstract:

When collecting body measurements of meat sheep using non-contact methods, the smooth texture of the sheep’ s skin and their constantly changing postures can lead to difficulties in locating feature points, inconsistencies in measurement reference points, and the loss of surface detail information. In response, an automated method for measuring the body dimensions of meat sheep was proposed based on a multi-module enhancement strategy. Datasets comprising 5 000 images each of sheep viewed from the side and from above were established. Using YOLO 11n-Pose as the base model, the keypoint detection model PDRE-YOLO was constructed by integrating the following modules: Faster_ Block with PConv, DCNv4 with dynamic deformable convolutions, RepViTBlock, and EVCBlock applied to the spine, neck, and head detection regions, respectively. Using PDRE-YOLO, nine key points were extracted: shoulder break point, shoulder height point, chest depth point, chest height point, rump height point, ischial tuberosity point, hoof tip point, left shoulder point, and right shoulder point. Based on the pixel values of these key points and coordinate transformation algorithms, six body measurements were automatically calculated: body length, body height, chest depth, chest height, cross height, and chest width. Experimental results indicated that on the side-view dataset, the model achieved a keypoint detection accuracy of 99. 50% for the mAP @ 0. 5: 0. 95, representing a 1. 8 percentage point improvement over the original YOLO 11n Pose model; detection accuracy reached 93. 60% , an increase of 3. 8 percentage points. On the top-view dataset, mAP@ 0. 5:0. 95 also achieved 99. 50% , an improvement of 1. 5 percentage points; with detection accuracy at 92. 60% , an increase of 2. 0 percentage points. The body measurements derived from coordinate-transformed keypoints were compared with manual measurements, yielding mean absolute errors (MAE)of 3. 94 cm, 3. 23 cm, 2. 04 cm, 1. 06 cm, 5. 47 cm, and 1. 42 cm, with average relative errors of 7. 63% , 5. 87% , 8. 75% , 4. 02% , 10. 22% , and 9. 68% , respectively.