The article is devoted to the actual problem of searching for contour images by given reference samples, due to the large amount of graphic scientific and technical information of the Contour type. Existing methods for image search do not allow taking into account the actual identity of images arising from affine transformations (stretching / compression, rotation) resulting from replication, which is important when analyzing graphic documentation. The paper proposes an alphabetical representation of contour images, which made it possible to construct a complex of affine-invariant equivalence relations. The relations built up served as the basis for the development of a set of selection procedures that can be considered as the initial stage of image search. The procedures included in the complex sequentially implement the reduction of the search scope up to a small set of contour images, allowing for the possibility of visual analysis. The complex is programmatically implemented in Python using the CUDA (NVIDIA GPU) software and hardware architecture for parallel computing.