基于模型剪枝的有色金属破碎料机器视觉分选方法优化  

Optimization of machine vision sorting method for nonferrous metal crushed aggregates based on model pruning

作　　者：董寰宇 秦训鹏[1] 丁吉祥 DONG Huanyu;QIN Xunpeng;DING Jixiang(School of Automotive Engineering,Wuhan University of Technology,Wuhan 430070)

机构地区：[1]武汉理工大学汽车工程学院,湖北武汉430070

出　　处：《机械设计》2023年第11期8-14,共7页Journal of Machine Design

基　　金：湖北省技术创新专项重大项目(2019AAA075)。

摘　　要：对破碎料图像进行准确、快速识别是实现有色金属破碎料实时分选的基础,文中基于YOLOv3模型,以BN层缩放因子为衡量指标对模型进行剪枝,剪切对运算结果没有影响的冗余通道,减少了卷积神经网络的参数量及计算量,实现了模型压缩。试验在生产现场采集的金属破碎料图像上进行,采用F1-Score、平均精度均值、浮点运算数、网络参数量和FPS作为评价指标,定量研究了模型压缩率对模型性能的影响。最终优化后,模型平均精度均值上升至97.1%,F1-Score为96.8%,参数量实现了70.4%的压缩率,浮点运算数降为原模型的44.5%,FPS上升40.4%,减少了模型运行计算量消耗和内存占用,加快了计算速度,能更好地满足工业生产中有色金属破碎料分选的需求。Accurate and fast recognition of fragment images is fundamental to realize real-time sorting of nonferrous metal crushed aggregates.In this article,based on the YOLOv3 model,efforts are made to prune the model with the BN layer scaling factor as the measurement index and prune the redundant channels that have no impact on the calculation results,thus reducing the amount of parameters and calculation of the convolution neural network,as well as realizing model compression.The experiment is carried out on the images of nonferrous metal crushed aggregates collected in the production site;With F1-Score,mAP,FLOPs,the network parameters and FPS as the evaluation indexes,the influence of the model's compression rate on detection accuracy is comparatively explored.Finally,mAP of the optimized model increases to 97.1%,F1-Score is 96.8%,the parameters enjoy the compression ratio of 70.4%,FLOPs is 44.5%that of the original model,and FPS increases by 40.4%.This reduces computational and memory consumption of model operation,and accelerates the speed of calculation,thus better meeting the needs of sorting of nonferrous metal crushed aggregates in industrial production.

关 键 词：深度学习 模型剪枝 有色金属 图像处理 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]