The method described represents an attractive compromise between the use of a single filter for a number of image distortions and one filter for each image or distortion of an image. The goal of this work is the generation of a number of filters, each of them being able to recognize a number of distortions. One of the main problems of the filter design is its high sensitivity to internal noise of the system, optical aberrations, etc. In this work, the most common case of image distortion invariance has been considered together with system noise invariance. The simulation results indicate the absence of a false alarm and good identification. The filter generation is based on a learning process in an electro-optical pattern recognition system. The genetic algorithm serves as an optimization method. A binary filter has been selected as the spatial filter. The comparison between a traditional matched spatial complex filter and this adaptive binary filter performance indicates the significant benefits of the latter. Statistical tools were used to estimate and compare the significance of the difference between the output average of a rejection and recognition image sets.