基于深度学习技术生猪图像目标检测算法的应用研究

Application of Pig Image Target Detection Algorithm Based on Deep Learning Technology

将基于深度学习的图像目标检测技术引入到养殖个体图像目标检测,可以提高养殖视频图像智能分析技术,提高科学养殖能力。试验将深度学习的YOLO V3算法应用到生猪图像目标检测,结合畜牧养殖实际情况,进行了类别选择、遮挡物处理和图像增强等设计,实现了基于深度学习技术的生猪图像目标检测算法。利用该算法对采集的生猪个体图像数据进行训练、验证和测试,对测试图像目标检测漏检率约6%,错检率约1%,精度较高;同时也与其他深度学习目标检测算法进行了对比和分析,测试结果反馈检测精度良好,检测速度较快,对比Fast R-CNN深度学习目标检测算法,mAP-50提高了7%~8%,检测速度提高了约5倍。与SSD算法比较,mAP-50指标和检测速度相当,但是由于YOLO V3算法网络模型比SSD算法简洁,算法移植兼容性更高。研究与试验结果表明,YOLO V3算法检测速度快,适合畜牧养殖图像智能识别工程化目标检测的要求。

The deep learning based image target detection technology was introduced into the target detection of individual breeding images,which can improve the intelligent analysis technology of breeding video images and enhance scientific breeding capabilities.YOLO V3 algorithm was applied to pig detection in videos.Based on the actual situation of animal husbandry,we designed category selection,occlusion processing and image enhancement,and implemented a pig image target detec⁃tion algorithm based on deep learning technology.The algorithm proposed was used to train,verify and test the collected pig image data.The missed detection rate of YOLO V3 was about 6%,and the error detection rate was about 1%.The results show that the algorithm has good detection accuracy and fast detection speed.On our dataset,we found that mAP-50 of our algorithm was improved by 7%-8%and detection speed was increased by about 5 times compared with Fast r-cnn.mAP-50 and detection speed of our algorithm are similar to SSD.However,because the Yolo V3 algorithm is sim⁃pler than SSD algorithm,the algorithm migration compatibility is higher.The results show that YOLO V3 algorithm has a fast detection speed and is suitable for the requirements of intelligent recognition engineering target detection in animal husbandry images.