Comparison of Analog and Discrete Hopfield Neural Network for Image Change Detection

This paper proposes an analog Hopfield neural network (HNN) for automatic image change detection problem between two images. This optimization relaxation approach differs from other techniques in that it provides the strength of the change rather than assigning binary labels (changed/unchanged) to each pixel. A difference image is obtained by subtracting pixel by pixel both images. The network topology is built so that each pixel in the difference image is a node in the network. Each node is characterized by its state, which determines if a pixel has changed or unchanged. An energy function is derived, so that the network converges to stable state. The main drawback of existing binary labeling approaches is that pixels are labeled according to the information supplied by its neighbors, where its self information is ignored. The main contribution of the analog Hopfield’s model is that it allows a tradeoff between the influence of a pixel’s neighborhood and its own criterion. This is mapped under the energy function to be minimized. Also a comparison between analog and discrete HNN shows similar Percentage of Correct Classification and Yule values. However, the analog counterpart describes the degree of change by embedding both Spatial-Contextual Information and Self-Data Information.