基于深度学习的笼养蛋鸡行为实时检测方法

针对蛋鸡养殖中,传统蛋鸡行为检测操作复杂、分类单一、实时性差的问题,提出一种基于深度学习的轻量型蛋鸡行为检测算法TD-YOLOV3。该检测算法以YOLOV3为基础网络结构,对其进行网络结构压缩,获得轻量型T-YOLOV3网络结构,用以提高系统检测速度;将第一个多尺度预测中的残差模块替换为Dense block,并在网络结构中的第Convolution 5,Convolution 7,Convolution 10,Convolution 12的卷积层之后添加NIN网络中的MLP结构,用以提高检测精度;采用基于K-means算法的聚类维度优化和训练策略优化对本研究的数据集进行训练和测试。试验结果表明,本研究提出的TD-YOLOV3检测算法的平均精准度均值89.26%,检测速度为33帧/s,参数量为55 MB;在同一硬件水平下与YOLOV3和T-YOLOV3相比,TD-YOLOV3在检测速度、精度等方面的综合性能最优,更适用于笼养蛋鸡行为的实时自动检测。

Real-time detection method for cage laying hens' behavior based on deep learning

Aiming at the problems in traditional egg hen behavior detection,single classification and poor real-time performance in laying hens,a lightweight hen behavior detection algorithm TD-YOLOV3 based on deep learning was proposed.The detection algorithm uses YOLOV3 as the basis network structure and compresses the network structure to obtain a lightweight T-YOLOV3 network structure to improve the system detection speed;The residual module in the first multi-scale prediction was replaced with Dense block and add the MLP in the NIN network after the convolutional layer of the Convolution 5,Convolution 7,Convolution 10 and Convolution 13 in the network structure to improve detection accuracy;The data set of this study was trained and tested by clustering dimension optimization and training strategy optimization based on K-means algorithm.The test results showed that the average precision of the TD-YOLOV3 detection algorithm proposed in this study was 89.26%,the detection speed was 33 frames per second,and the parameter quantity was 55 MB.Compared with YOLOV3 and T-YOLOV3 at the same hardware level,TD-YOLOV3 has the best overall performance in terms of detection speed and accuracy,and is more suitable for real-time automatic detection of cage laying behavior.