Improving the image recognition capability of Hopfield neural networks

Hopfield neural networks can be used for image recognition when only a partial image is available. However, the image recognition process is very sensitive to the position of the input; shifting the image by only one pixel can cause the network to fail to find a matching exemplar. The authors present a technique for modifying the input image so that an ordinary Hopfield neural network will recognize a shifted image. This technique makes use of the image. The authors run an experiment with random bitmap images to determine how accurately a Hopfield neural network can recognize shifted and blurred images. The results indicate that the neural network can recognize shifted images only if they are modified