Category-aware Edge Detection Model Design Based On Deep Learning

Edge detection is always a fundamental research problem in image processing,as well as a technical basis for many other research issues.The existing traditional methods do not have a specific edge screening function,while the deep learning based method can perform specific edge screening and classification tasks,but there is a problem of low accuracy.For some highly demanding tasks,such as fine segmentation of specific types of cells in medical images,existing algorithms do not meet the needs.To this end,this paper proposes a more accurate category-aware edge detection algorithm based on deep learning.After researching and analyzing the characteristics of the existing model and combining with the improvement ideas,a new model is proposed.In the process of model design,the three aspects of feature fusion strategy,supervised learning strategy and loss function design are considered.For feature fusion,in order to compensate for the loss of detail features due to network deepening,a multi-level feature fusion method is used to fuse the hierarchical features of the first three stages of the model into the final multi-class feature.In the aspect of supervised learning strategy design,a deep supervised learning method was adopted.In the first three stages of the model,a label-scale adaptive binary edge supervision was used,and a new alternate supervision method was used in the final multi-classification stage.In order to achieve the desired supervised effect,the loss function is specifically designed.The binary edge and the first multi-class edge are supervised using the reweighted cross-entropy loss,while the final multi-class result is supervised by normal cross-entropy loss.The experimental proves that the accuracy of the model is greatly improved by this design.In order to verify the validity of the model,the two public data sets SBD and Cityscape are selected for data verification,and the CASENet model is selected for the comparison model.The test results show that the average ODS F-measure of the model on the SBD dataset is increased by 3% and by 6% on the Cityscape dataset.This verifies the effectiveness of the model design strategy in this paper.