使用候选框进行全卷积网络修正的目标分割算法

Object segmentation algorithm modified by candidate box for fully convolution network

由于反卷积和上池化操作的存在,传统全卷积网络在解码阶段常常会丢失目标位置信息,降低图像的分割精度。针对这种情况,提出基于候选框网络对全卷积网络的输出进行缺陷位置微调的液晶面板缺陷分割算法。算法基于ResNet-101网络搭建全卷积主干网络,此构建2个分支,候选框生成网络和反卷积网络。在反卷积网络的输出层中使用多通道分类损失函数,输出关于每种缺陷的类别分割图。同时利用候选框网络产生高置信度的目标框,以此框对反卷积网络输出的类别分割图进行逐通道修正,使用修正后的多通道缺陷类别分割图进行逐像素分类,得到最终分割结果。实验结果表明,该算法对液晶面板缺陷的分割取得了7.5%的精度提升,边缘分割更加精细化。

Due to the existence of deconvolution and up-pooling operations,traditional full-convolution networks often lose target location information during decoding,reducing image segmentation accuracy.Aiming at this situation,a liquid crystal panel defect segmentation algorithm based on the candidate frame network for fine-tuning the defect position of the output of the full convolution network is proposed.The algorithm first builds a full-convolution backbone network based on the ResNet-101 network,and then builds two branches based on this,one is the region proposal generation network,and the other is the deconvolution network.A multi-channel classification loss function is used in the output layer of the deconvolution network to output a class segmentation map for each defect.At the same time,the region proposal network is used to generate a high-confidence object proposal,and then the frame segmentation map outputted by the deconvolution network is corrected channel by channel.Finally,the modified multi-channel defect class segmentation map is used for pixel-by-pixel classification to obtain the final segmented results.The experimental results show that the algorithm achieves a 6.5%accuracy improvement in the segmentation of liquid crystal panel defects,and the edge segmentation is more refined.