The authors consider vision from a neural-network (NN) point of view. After describing the structure of the human retina and its neurons, a general model of NNs is presented. For color, patterns, and motion detection, a three-layer NN is detailed. Further, three variable-threshold NNs (color identification) are introduced. The detection of motion is also done in this second layer using already preprocessed information represented by on-off, on, and off signals, with the same meaning as the corresponding signals from the human ganglion cells; four analog classifiers are used for directions (up, down, left, right). The results of a program used to simulate the proposed NN are discussed, as well as a theoretical VLSI implementation. It is shown that a complete graph having n nodes can be laid out in an almost square area of O(n2), with O(n) wire length.