基于注意力机制和可变形卷积的鸡只图像实例分割提取

为提高鸡只个体轮廓分割提取的精度和准确度，实现基于机器视觉技术的鸡只行为、健康、福利状态监测等精准畜牧业管理，保证相关监测技术及决策的可靠性，针对叠层笼养环境下肉鸡图像的实例分割和轮廓提取问题，提出一种优化的基于Mask R-CNN框架的实例分割方法，构建了一种鸡只图像分割和轮廓提取网络，对鸡群图像进行分割，从而实现鸡只个体轮廓的提取。该网络以注意力机制、可变形卷积的41层深度残差网络（ResNet）和特征金字塔网络（Feature pyramid networks, FPN）相融合为主干网络，提取图像特征，并经区域生成网络（Region proposal networks, RPN）提取感兴趣区域（ROI），最后通过头部网络完成鸡只目标的分类、分割和边框回归。鸡只图像分割试验表明，与Mask R-CNN网络相比，优化后网络模型精确率和精度均值分别从78.23%、84.48%提高到88.60%、90.37%，模型召回率为77.48%，可以实现鸡只轮廓的像素级分割。本研究可为鸡只福利状态和鸡只健康状况的实时监测提供技术支撑。

Instance Segmentation of Broiler Image Based on Attention Mechanism and Deformable Convolution

Segmentation and extraction of birds contour is the premise of precision livestock farming management, such as behavior, health, welfare status monitoring based on machine vision technology. The precision and accuracy of image segmentation directly affect the reliability of relevant monitoring technology and decision-making. An instance segmentation approach based on Mask R-CNN deep learning framework was proposed to solve broiler instance segmentation and contour extraction problems in stacked-cage henhouse. Furthermore, a broiler image segmentation and contour extraction network was constructed to segment broiler images and realize birds individual contour extraction. In this network, totally 41 layers deep residual network (ResNet) based on attention mechanism and deformable convolution was integrated with feature pyramid networks (FPN) as the backbone network to extract the image features, and regions of interest were extracted by region proposal networks. Finally, target classification, segmentation and box regression were realized through network heads. Broiler image segmentation experiment showed that compared with Mask R-CNN network, the average precision and mean accuracy of the optimized network were improved from 78.23% and 84.48% to 88.60% and 90.37%, respectively, and the recall rate of the model was 77.48%, which can realize the pixel level segmentation of chicken contour. The research result can provide technical support for the real-time monitoring of birds welfare and health status.