基于DCT-YOLO的轻量化奶牛爬跨行为检测方法

Lightweight Cow Mounting Behavior Detection Method Based on DCT-YOLO

摘要:

针对现有接触式奶牛发情检测装置成本高，易造成奶牛应激反应，部分基于深度学习的检测方法存在复杂环境影响下识别精度不高、模型复杂度高及部署困难等问题，基于YOLO v8n模型提出了DCT-YOLO的轻量化奶牛爬跨行为与发情奶牛检测模型。首先，骨干部分的特征提取与下采样采用YOLO v7的MPConv(MaxPool-Conv)模块，提升模型对远景小目标奶牛爬跨行为的识别能力；其次，在检测头部分采用动态任务对齐检测头(Task align dynamic detection head, TADDH)，融合爬跨行为与发情奶牛间的特征关联，达到通过爬跨行为这一显著特征增强网络对发情奶牛个体关注度的目的；上采样部分采用内容感知重组模块(Content-aware reassembly of features, CARAFE)，通过跨维度交互进行发情奶牛特征增强。为了验证模型性能，标注了2239幅图像用于模型训练与测试。试验结果表明，DCT-YOLO模型精确率为94.8%，召回率为80.1%，平均精度均值(mAP@0.5)为87.5%，浮点计算量为8.5×109，参数量为2.08×106，检测速度为256.41f/s。与SSD、Faster R-CNN、YOLO v5n、YOLO v5s、YOLO v7-tiny、YOLO v8n和YOLO v8s目标检测模型相比，参数量分别降低91.19%、98.48%、16.93%、77.18%、65.41%、30.82%和81.31%，检测速度分别提高192.72、226.56、34.19、97.68、187.92、39.02、126.54f/s，平均精度均值(mAP@0.5)仅比YOLO v8s低1.9个百分点，比其他模型的最高值高0.2个百分点，结果表明，模型取得了检测精度与速度的良好平衡。综上，本研究可为奶牛爬跨行为识别与发情奶牛定位等任务提供关键信息支撑。

Abstract:

Accurate and effective cow estrus detection is the foundation for improving herd reproductive performance. Existing contact cow estrus detection devices are costly and prone to cause stress in cows, and some deep learning-based detection methods suffer from poor detection accuracy under the influence of complex environments and deployment difficulties due to high model complexity. Therefore, a lightweight DCT-YOLO model for cow climbing spanning behavior and estrus cow detection was proposed based on the YOLO v8n model for improvement. Firstly, the MPConv structure was adopted for the feature extraction and subsampled of the Backbone part to improve the recognition ability of the model for small-target cow mounting behavior. Secondly, the detection head adopted a dynamic TADDH, which fused the feature association between estrous cows and mounting behavior, and improved the network’s focus on individual estrous cows through the distinctive feature of mounting behavior. Finally, the interpolating part adopted CARAFE, which enhanced the features of estrus cows through cross-dimensional interactions. To validate the performance of the model, totally 2239 images were labeled for model training and testing. The experimental results showed that DCT-YOLO model had a precision of 94.8%, a recall of 80.1%, a mean average precision (mAP@0.5) of 87.5%, floating-point operations (FLOPs) of 8.5×109, params of 2.08×106, and a detection speed of 256.41f/s. Compared with SSD, Faster R-CNN, YOLO v5n, YOLO v5s, YOLO v7-tiny, YOLO v8n and YOLO v8s target detection networks, the number of parameters was reduced by 91.19%, 98.48%, 16.93%, 77.18%, 65.41%, 30.82% and 81.31%, and the detection speed was respectively improved by 192.72f/s, 226.56f/s, 34.19f/s, 97.68f/s, 187.92f/s, 39.02f/s and 126.54f/s, respectively, mAP@0.5 was only 1.9 percentage points lower than that of YOLO v8s and 0.2 percentage points higher than that of the best of the other models and the results showed that the model achieved a good balance between detection accuracy and speed. In summary, the model was lightweight, real-time accurate and robust, and it can provide important information support for tasks such as cow mounting behavior detection and estrus cow localization.