引入注意力机制的蚕茧识别算法  

Cocoon recognition algorithm based on attention mechanism

作　　者：陈国栋[1] Chen Guodong(Soochow University,Suzhou 215000,China)

机构地区：[1]苏州大学,江苏苏州215000

出　　处：《纺织报告》2020年第11期23-25,28,共4页

摘　　要：当前,蚕茧识别方法大多使用费时费力的人工检测,利用Faster R-CNN目标检测框架进行自动识别来节约成本。Faster R-CNN在分类回归任务前对每张蚕茧图片进行特征提取,在提取特征过程中,下采样的pooling操作导致部分特征丢失,对蚕茧类等尺寸较小的物体检测率较低,存在容易漏检、误检等缺陷。针对缺陷提出了加入注意力机制的Faster R-CNN算法,主要思想是使用空洞卷积代替下采样过程的pooling操作,对输出的特征图经过注意力机制网络进行微调,减少高层次特征损失。蚕茧数据集上的实验表明,修改后的算法平均mAP提高了2.5%。Most of the current cocoon identification methods use time-consuming and laborious manual detection.Faster R-CNN target detection framework can be used for automatic identification to save cost.Feature extraction is conducted for each cocoon image before the classification regression task.During the feature extraction process,partial feature loss is caused by the pooling operation of the lower sampling,and the detection rate of objects of small size such as silkworm cocoons is low,with defects such as easy missing detection and false detection.The main idea is to replace the pooling operation in the under-sampling process with empty convolution,and meanwhile,the output feature mAP is fine-tuning through the attention mechanism network to reduce the high-level feature loss.The experiment on cocoon data set shows that the modified algorithm improves the average mAP by 2.5%.

关 键 词：蚕茧识别 Faster R-CNN 注意力机制 空洞卷积 

分 类 号：TP312[自动化与计算机技术—计算机软件与理论]