改进的Faster R-CNN网络预测尺度参数  

Improved Faster R-CNN algorithm for object parameter prediction

作　　者：王婷婷 苍岩[1] 毕晓君[1] 何恒翔 WANG Tingting;CANG Yan;BI Xiaojun;HE Hengxiang(College of Information and Communication Engineering, Harbin Engineering University, Harbin 150001, China;School of Electronics and Information Engineering, Harbin Institute of Technology, Harbin 150006, China)

机构地区：[1]哈尔滨工程大学信息与通信工程学院,黑龙江哈尔滨150001 [2]哈尔滨工业大学电子与信息工程学院,黑龙江哈尔滨150006

出　　处：《哈尔滨工程大学学报》2021年第3期426-432,共7页Journal of Harbin Engineering University

基　　金：国家自然科学基金项目(51779050).

摘　　要：为了研究基于深度学习的多任务分类和回归问题,本文改进了Faster RCNN网络,提出了同时实现分类和回归分支的深层神经网络。以Faster R-CNN算法为基础,通过增加网络分支的方式,将分类和回归网络并入同一个网络,实现端对端的目标识别、定位和参数测量多个任务并行处理。使用整体交替训练的方式优化网络参数。基于智慧农业应用背景,利用生猪的体型特征与体重参数之间的相关性,实现了通过生猪体型图像预测生猪体重参数的任务。实际场景中利用该网络完成了生猪体重参数预测,预测误差0.374%。In this paper,a Faster R-CNN network is modified and a deep neural network is proposed to predict object classification and parameter regression calculation by the deep learning algorithm.The improved network realizes classification and regression at the same time.Based on the Faster R-CNN algorithm,by adding a regressive branch,the classification and regressive networks are integrated into an end-to-end network to realize object recognition,location,and parameter prediction paralleling.The alternate training method optimizes the parameters of the network.In addition,based on the application of artificial intelligence in agriculture,by the correlation between the shape and weight of pigs,the weighing prediction by the proposed network is realized by pig shape image,with a prediction error of 0.374%.

关 键 词：深度学习 Faster R-CNN算法 无接触测重 目标参数测量 交替训练 分类 回归 体重预测 

分 类 号：TP391.9[自动化与计算机技术—计算机应用技术]