NOISE ROBUST COLOR EDGE DETECTION USING FEATURE FUSION WITH MULTI-LAYER PERCEPTRON

Edge detection algorithms have achieved excellent results on clean images. However, both traditional methods with hand-crafted features and popular deep learning approaches with auto-learned features struggle with images corrupted by noise. To enhance the robustness against various types of noise, this paper proposes a method which combines features from both spatial and range domains for edge detection. Spatial features are extracted using anisotropic directional derivatives filters, while range features are extracted by anisotropic morphological derivatives. Leveraging the complementary nature of multi-domain information for noise suppression, a Multi-layer Perceptron is employed to learn the non-linear relationship between the fused features and the edge probabilities, thereby creating an edge detection model which is robust to noise. Experimental evaluation on a color edge dataset, benchmarked against state-of-the-art methods, validates the effectiveness and robustness of the proposed method.