基于多尺度感知的高密度猪只计数网络研究

猪只盘点是生猪规模化养殖和管理中的重要环节，人工计数方法费时、费力，在大数据量的猪只盘点中容易出错。本文使用多尺度感知网络对高密度猪群图像中的猪只进行计数。通过对人群计数网络CSRNet的改进，得到猪只计数网络（Pig counting net, PCN），PCN采用VGG16作为前端网络提取特征，中间层采用空间金字塔（Spatial pyramid）结构对图像中的多尺度信息进行提取与融合，后端网络采用改进的膨胀卷积网络。PCN增加了多尺度感知结构、扩大了后端网络感受野，通过感知多尺度特征得到预测密度图，预测密度图反映了猪只空间分布，通过对密度图积分实现了猪只数量的估计。结果表明，在平均猪只数为 40.71的测试集图像上，PCN的计数准确率优于人群计数网络 MCNN、CSRNet和改进Counting CNN 的猪只计数网络，MAE和RMSE 分别为1.74和 2.28，表现出较高的准确性和鲁棒性；单幅图像平均识别时间为0.108s，满足实时处理要求。

High-density Pig Counting Net Based on Multi-scale Aware

Pig inventory is an important part of large-scale breeding and management of live pigs. Manual counting methods are more time-consuming and laborious, especially in pig inventory with large amounts of data. How to count high-density pig herd images with machine vision is still a difficult problem to be solved urgently. A multi-scale aware counting network was used to count pigs in high-density pig herd images. Based on the crowd counting network CSRNet, the pig counting network of pig counting net(PCN)was proposed. VGG16 was used as the front-end network to extract features, the spatial pyramid structure was used, and this structure can extract and fusion multi-scale information in the image, the back-end network used an improved dilated convolutional network. PCN added a multi-scale aware structure, expanded the back-end network receptive field, and can obtain a predicted density map by sensing multi-scale features, the predicted density map reflectedthe spatial distribution of pigs, then by integrating the density map, the number of pigs can be accurately calculated.The results showed that on the test set image with an average number of pigs of 40.71, the accuracy of PCN was better than that of the crowd counting net MCNN, CSRNet and the pig counting net that modified Counting CNN, the mean absolute error (MAE) and the root mean square error (RMSE) were 1.74 and 2.28, respectively,the lower error showed that PCN had better accuracy and robustness.The average recognition time of a single image of the final model was 0.108s, which met the real-time processing requirements of the algorithm.The method provided a research idea for the automatic inventory of high-density group raising pigs.