Color edge detection by learning classification network with anisotropic directional derivative matrices

摘要

In this paper, a neural network-based color edge detector is constructed by learning a classifier using anisotropic directional derivative (ANDD) matrices of a color image as input. The training stage on a color edge dataset with ground truth (GT) edges includes calculation of ANDD matrices, generation of feature matrices, and training a classifier. For each training image, a set of ANDD matrices are calculated from the ANDDs with different parameter setups for training and from which a set of the color edge strength maps (CESMs) are extracted by the singular vector decomposition. The CESMs and the GTs on edges of the image are combined into a feature matrix for training. Using the feature matrices of all the training images as input, a classification neural network is trained and it outputs the probability of a pixel to be an edge pixel. In the detection stage, for a color image, its ANDD matrices, CESMs, and the color edge direction maps (CEDMs) are first computed and then the CESMs are input into the classification neural network to obtain the edge probability map (EPM) of the image. Finally, the non-maximum suppression and hysteresis thresholding are applied to the EPM and CEDMs to generate the binary edge map. The proposed detector attains better performance than the existing gradient-based detectors and is competitive with learning-based detectors on three commonly-used color image datasets for edge and contour detection.