Visual Cross-Image Fusion Using Deep Neural Networks for Image Edge Detection

Abstract

Edge detection is a fundamental computer vision problem and has wide applications, Convolutional neural networks (CNN) has been a good fundamental component in many image edge detection systems. But the edge detection accuracy from the detection and extraction of edge features by CNN is still below human visual perception, the edges of lines are fuzzy and the process is time-consuming. Based on the convolutional architecture for fast feature embedding (Caffe) framework and Visual Geometry Group (VGG16) template, a new accurate edge detector is presented in this paper by using visual cross fusion (VCF) network to detect features first. Our VCF model extracts the features of the multi-level hierarchy by parameter dimension reduction and cross-fusion of the fully connected layers respectively to achieve the end-to-end image edge detection. Second, for further improvement, a custom grading weighted cross-entropy loss function is to maximize the use of image pixels set and organize the disputed pixels rationally. Finally, the cross-network fusion layer is used to refine the edge features of images. The experimental results show that the VCF algorithm achieves an ODSF-measure as 0.808 on the Berkeley Segmentation Data Set (BSDS500) classic dataset, which is 2.5% higher than the Holistically-Nested Edge Detection (HED) algorithm, while maintaining fast 30+ FPS. On the NYU Depth Dataset V2 (NYUD V2) dataset, the ODSF-measure is 0.783, improved by 5.66% compared with the HED algorithm, and the FPS is increased by 10+. Therefore, the VCF algorithm is closer to the average human visual perception.

Keywords

• Edge detection
• deep learning
• convolutional neural network
• VGG16
• loss function