When classifying images, in order to reduce the dimension of the feature space, it is necessary to convolve images to select a subset of pixels sufficient to solve the problem. An image convolution method based on the correlation of the pixels of images of the training sample with a teacher with an approximated function used for classification is proposed. An example of image convolution used to classify handwritten digits by evaluating the posterior probabilities of ten image classes using Anderson's discriminant function is given.