The Supervised Research On CNN Image Edge Detection Algorithm In Scotopic Vision Environment

The image acquired by image acquisition equipment in scotopic vision environment is called scotopic vision image,which has the characteristics of low contrast and low brightness.When the existed edge detection algorithms extract edges of these kinds of images,the problem of increasing false edges and losing part of edges was easy to occur.To improve the above problems,an improved Convolutional Neural Networks edge detection Dexi Ned(Dense Extreme Inception Network for Edge Detection)algorithm was studied in this thesis.The algorithm can directly extract the most effective edges from scotopic vision images and improve the edge detection accuracy of normal light images to a small extent.The main work of the thesis includes:1.Dexi Ned algorithm performs well in edge detection of normal light images,but there are many false edges appeared when scotopic vision images are detected.An improved Dexi Ned model was proposed in this thesis,which retained the backbone network of the Dexi Ned model,by adding multiple convolutional layers and residual units in the appropriate position of the Dexi Ned model,the ability of learning edge features during model training was enhanced.The improved Dexi Ned model not only eliminated most of the false edges generated by the Dexi Ned model in the scotopic vision image,but also extracted most of the effective edges directly from the scotopic vision image.The results of the edge data set(Barcelona Images for Perceptual Edge Detection,BIPED)showed that the fixed contour threshold,the optimal threshold of each image and the average accuracy of the model were improved to 0.879,0.885 and0.899,respectively.However,when the model is applied directly to scotopic vision image,part of the edge information will still be lost.2.In order to improve the situation that the improved Dexi Ned model lost the edge in the scotopic vision image and improved the edge detection accuracy of the model for the scotopic vision image.The edge detection results of scotopic vision images were optimized based on data sets in this thesis.Inspired by the principle that the deep learning model is a data-driven way to achieve specific tasks,it was considered to strengthen supervision from the source of the data set.However,there is no public available data set for edge annotation of scotopic vision images at present.Thus combining the feature of scotopic vision image in the field of digital image,based on BIPED data sets,this thesis build suitable for the training set of scotopic visual image edge detection from RGB,YUV,YCBCR color spaces respectively,improving the edge detection precision of the model.In the collection of scotopic vision images test set and the test set of LOL(Low Light Dataset)images meeting the conditions(the background gray level is less than 47),the results show that they performed well in the edge detection model based on the training set constructed by RGB color space best.Among them,the Mean Square Error of the scotopic vision image test set decreased to 0.006321,the Peak Signal-to-Noise Ratio and the Structural Similarity increased to 8.936 and 0.838,and the values of three evaluation indexes of the LOL image test set were 0.00768,8.474 and 0.846,respectively.The edge detection model studied in this thesis makes up for the lack of contrast resolution of human vision to some extent and can observe the effective edge information of scotopic vision image with a lower cost.