Edge-backpropagation for noisy logo recognition

摘要

In this paper, we propose a new approach to improve the performance of multilayer perceptrons operating as autoassociators to classify graphical items in presence of spot noise on the image. The improvement is obtained by introducing a weighed norm instead of using the Euclidean norm to measure the input–output accuracy of the neural network. The weights used in the computation depend on the gradient of the image so as to give less importance to uniform colour regions, like the spots. A modified learning algorithm (edge-backpropagation) is derived from the classical backpropagation by considering the new weighed error function. We report a set of experimental results on a database of 134 company logos corrupted by artificial noise which show the effectiveness of the proposed approach.