The Supervised CNN Image Edge Detection Algorithm in Scotopic Vision Environment

Scotopic vision environment images are characterized by low contrast and some details hidden in the image background, that causes human eyes are hard to detect and brings difficulties to the subsequent application of computer vision tasks. In order to solve the problem that many false edges are generated when DexiNed (Dense Extreme Inception Network for Edge Detection) model detects scotopic vision images, an improved DexiNed edge detection model was proposed in this paper. The improved edge detection model retained the backbone network of the DexiNed model. By adding the convolution layers and residual units in the appropriate position of the DexiNed model, the model can eliminate most of the false edges generated by the DexiNed model in the scotopic vision images better. In order to further improve the edge detection accuracy of scotopic vision image by the improved DexiNed model, this paper builds scotopic vision image training set based on edge annotation data set BIPED (Barcelona Images for Perceptual Edge Detection) from RGB and YUV color space respectively. And scotopic vision image test dataset results showed that the effect of scotopic vision image based on RGB color space to have better performance, because edge continuity was better and the edge detection model of MSE (mean square error) index dropped, PSNR (peak signal to noise ratio) and SSIM (structural similarity) index raised.