图像分割算法在肉鸡深度图集上的研究

Research on Image Segmentation Algorithms on Broiler Depth Atlas

【目的】针对在复杂环境背景中难以识别分割多只肉鸡的问题,探讨基于深度学习实现对多只肉鸡深度图像分割的方法。【方法】利用深度相机,通过不同的拍摄角度(俯视、正视、侧视)在自然环境下采集肉鸡不同姿势(站立、俯卧、抬头、低头等)形态的深度图像,并使用CVAT标注软件对深度图像进行精确标注,建立肉鸡深度图数据集(含4 058张深度图像)。利用FCN、U-Net、PSPNet、DeepLab和Mask R-CNN等5种神经网络实现肉鸡深度图像的识别与分割,根据测试集得到预测结果,比较与评估不同模型的性能,实现对肉鸡深度图像的识别与分割。【结果】基于Mask R-CNN神经网络模型的识别分割准确率为98.96%,召回率为97.78%,调和平均数为95.03%,交并比为94.69%,4个指标值均为5个模型中的最优值。【结论】基于Mask R-CNN神经网络的算法简单快速,且能准确实现肉鸡的自动识别与分割,对肉鸡遮挡有较佳的鲁棒性,基本可以满足养殖场鸡群均匀度预测的识别分割要求。促进了计算机视觉在现代农业的应用,可为鸡群计数、鸡群均匀度预测以及肉鸡福利饲养等鸡场作业提供理论和实践基础。

【Objective】Aiming at the difficulty in recognizing and segmenting multiple broilers in complex environment, a segmentation method for depth map of multiple broilers based on deep learning was explored.【Method】By using the depth camera, the depth map of broilers in different postures(standing, prone, looking up, looking down, etc.)were collected in the natural environment through different shooting angles(top, front and side), and the depth map were accurately marked by CVAT labeling software. A broiler depth map dataset was established, with a total of 4 058 depth maps. Five neural networks, including FCN, U-NET, PSPNet,DeepLab and Mask R-CNN, were used to recognize and segment broiler depth maps. Based on the predicted results of test sets, the performance of different models were compared and evaluated to realize the recognition and segmentation of broiler depth maps.【Result】The recognition and segmentation accuracy of Mask R-CNN neural network model is 98.96%, the recall rate is 97.78%, the F1 score is 95.03%, and the intersection-overunion is 94.69%, all of which are the optimal values of the five models.【Conclusion】The algorithm based on Mask R-CNN is simple and fast, and it can realize the automatic recognition and segmentation of broilers accurately and has good robustness to the shielding of broilers, which can basically meet the recognition and segmentation requirements for the prediction of the evenness of chicken flocks in the chicken farm. It promotes the application of computer vision in modern agriculture, and provides theoretical and practical bases for chicken farm operations such as flock counting, flock evenness prediction and welfare breeding of broilers.