Neural networks for robust image feature classification: a comparative study

We propose a simple and powerful feature extractor using neural networks. This feature extractor is trained to detect features such as lines, corners, junctions in images. Different feature models are generated based on discontinuity in intensity values and the orientation of the boundary in the pixel neighborhood. Locating feature points in the image is carried out in two steps by considering an n×n window as a processing unit. At the first step, a covariance technique is used to calculate rotation-invariant descriptors, which represent discontinuities for edge types. At the second step, a multilayer feedforward neural network, trained with the invariant feature descriptors, is used to classify the centre pixel into one of the possible features. Experimental results using the proposed method are compared with Marr-Hildreth edge operator results to show the effectiveness of the proposed method