采用改进CenterNet模型检测群养生猪目标

为实现对群养环境下生猪个体目标快速精准的检测，该研究提出了一种针对群养生猪的改进型目标检测网络MF-CenterNet（MobileNet-FPN-CenterNet）模型，为确保目标检测的精确度，该模型首先以无锚式的CenterNet为基础结构，通过引入轻量级的MobileNet网络作为模型特征提取网络，以降低模型大小和提高检测速度，同时加入特征金字塔结构FPN（Feature Pyramid Networks）以提高模型特征提取能力，在保证模型轻量化、实时性的同时，提高遮挡目标和小目标的检测精度。该研究以某商业猪场群养生猪录制视频作为数据源，采集视频帧1 683张，经图像增强后共得到6 732张图像。试验结果表明，MF-CenterNet模型大小仅为21 MB，满足边缘计算端的部署，同时对生猪目标检测平均精确度达到94.30%，检测速度达到69 帧/s，相较于Faster-RCNN、SSD、YOLOv3、YOLOv4目标检测网络模型，检测精度分别提高了6.39%、4.46%、6.01%、2.74%，检测速度分别提高了54、47、45、43 帧/s，相关结果表明了该研究所提出的改进型的轻量级MF-CenterNet模型，能够在满足目标检测实时性的同时提高了对群养生猪的检测精度，为生产现场端的群养生猪行为实时检测与分析提供了有效方法。

Detection of group-housed pigs based on improved CenterNet model

Abstract: Rapid and accurate detection of pigs has been critical to intelligent monitoring of health status within a group-housed breeding environment on large-scale farms. However, a large number of parameters make it difficult to achieve real-time performance in edge computing platforms for practical production. In this study, an improved CenterNet model (named MF-CenterNet) was proposed to detect pigs in group-housed breeding conditions, in order to improve the real-time performance of detection and the accuracy of localizing pigs with body occluded and small body size. An anchor-free CenterNet was also used to ensure the accuracy of detection, especially for the pig with body occluded. A lightweight MobileNet network was first introduced into the CenterNet (instead of ResNet50), as the backbone network of feature extraction for the smaller model size and higher detection speed. In addition, the feature pyramid structure (FPN) was then added to improve the ability of feature extraction for small pig objects. As such, the CenterNet was integrated with the MobileNet and FPN, named MF-CenterNet (i.e., MobileNet-FPN-CenterNet, MF-CenterNet). An image dataset of a private pig was collected to evaluate the performance of MF-CenterNet. All images were then captured from Jincheng Farm, Qiqihar City, Heilongjiang Province, China. Specifically, 1683 video frames were extracted from the video recording of pigs collected in the commercial pig farm, and 6732 images were obtained with the operation of the data argument. All pig objects within the images were then labeled with the labeling tool. The experimental results show that the size of the MF-CenterNet model was only 21MB, which satisfied the deployment of the model to an edge computing platform. The mean average precision (mAP) of pig detection was 94.30%, and the detection speed was up to 69 frames/s. The model of CenterNet integrated with MobileNetv2 achieved the best performance, in terms of accuracy, speed, and model size, where different versions of Mobile Net were combined. The CenterNet model integrated with the MobileNetv2 and FPN (MF-CenterNet) further improved the detection performance, indicating more robust in detecting the pig objects with small body size and body occluded. The improved MF-CenterNet greatly increased the mAP by 0.63 Percentage points , and the speed by 42 frames/s, while the size of the model was reduced by 104 MB, compared with the original CenterNet. Furthermore, the mAP detection was improved by 6.39, 4.46, 6.01, and 2.74 Percentage points , while, the detection speed was improved by 54, 47, 45, and 43 frames/s, respectively, compared with the common CNN-based object detection models, including Farster RCNN, SSD, YOLOV3, and YOLOV4 model. Consequently, the MF-CenterNet achieved the state-of-the-art mAP performance, higher detection speed, and the deployability of the model in a substantial manner. Therefore, this lightweight object detection model, MF-CenterNet, can meet the requirements of real-time, rapid, and high accuracy of detection on the group-housed pigs. The finding can also be expected to serve as a new way for real-time detection and prerequisite model in the behavior analysis of pigs during modern intensive production.