The Robustness of the Edge Detection on Noisy Natural Images with HED and PiDiNet Networks

Abstract

The edge detection performance of the HED and PiDiNet networks is compared on noised natural images. To research the robustness of edge detection with HED and PiDiNet networks, the technique of detecting the influence of the noise level on the CNN edge detector efficiency is proposed. At first, as a result of literature review, HED and PiDiNet are selected to detect edges on noisy images. Then types of image noise of interest to the researcher and the noise parameters which will be controlled by the researcher are determined. We considered white Gaussian noise, impulse noise, and speckle noise. Mathematical modeling of noisy images is performed. Next, the training and test sets from image datasets for edge detection is selected. After that, the researched networks are learned to detect edges on training set of images or weights of pre-trained networks are loaded. Further the measures of edge detection performance are selected, specifically, Precision, Recall, F1-score, and Pratt's Figure of Merit. Then images of test set are corrupted with controlled level of noise. The trained networks are applied to noisy images and the edge detection performance is evaluated depending on the noise level in the images. The obtained results are analyzed to generate the recommendations allowing to determine which CNN is better for edge detection when natural images affected by different level of white Gaussian noise, or impulse noise, or speckle noise. In particular, the HED network is generally preferred at high noise levels. The PiDiNet network is best used when the noise level is low.