Research on Edge Detection of U2NET Based on Differential Convolution
Edge detection plays an important role in the field of computer vision,which can be used in target detection,image segmentation and other tasks.U2NET is a deep learning model based on U-Net,which realizes high precision image segmentation through hierarchical feature extraction and context-aware feature fusion.This paper studies the application of U2NET in edge detection,and finds that the traditional convolution check is insensitive to gradient information in the convo-lution process.When the traditional convolution block is convolution,kernel optimization is ran-domly initialized,and the gradient information is not displayed and encoded,which makes it diffi-cult for it to focus on edge related features.Differential convolution replaces the original pixels in the local feature patch covered by the convolution kernel with pixel differences,encodes the useful pixel relations,and retains them in the convolution kernel during the training process.In this way,the problem that traditional CNN is not sensitive to capturing gradient information can be over-come.The results of experiments on self-made data sets and BSDS500 public data sets demon-strate the superior performance of differential convolution U2NET in edge detection tasks.In summary,the results of this paper show that the differential convolution based U2NET is an effi-cient model for matting and edge detection,which can be applied to various practical scenarios.
NETL
NSTL
万方数据
