Multi-Scale Auto-Encoder for Edge Detection

Abstract

A multi-scale encoder algorithm is proposed for image edge detection, which takes the auto-encoder as basic backbone structure. Three auto-encoders, each is responsible for processing an image of one scale, are organized together to perform image-to-image prediction by combining all multi-scale convolutional features. Taking the advantage of the multi-scale strategy and self-attention mechanism, the algorithm detects image edges from coarse to fine gradually, and succeeds in detecting edges missed easily in other works. Two types of skip connections designed in the model structure. One is connected between encoder and decoder within an auto-encoder, worked as the residual function, the other is between auto-encoders, providing multi-scale knowledge about the edges. The loss function is composed of cross-entropy function and Dice coefficient term which is used to handle imbalanced training data. Experimental results evaluated from BSDS500 and BIPED datasets show the algorithm achieves scores of 0.847, 0.858, 0.884 on ODS, OIS, AP indices and output high-quality edges.