Industrial printed matter image registration method based on deep learning and device thereof

Abstract

The invention discloses an industrial printed matter image registration method based on deep learning and a device thereof, and the method specifically comprises the steps: extracting slices with the same size from a template image and a to-be-registered image according to a corresponding cutting rule, obtaining a group of slices, carrying out the deep fusion of the group of slices, and obtaining an image slice pair, inputting the image slice pair into a registration network model for training to obtain a registered image slice; and rejecting edge parts of the registration image slices, and then performing slice alignment splicing to obtain a complete registration printed matter image. The framework of the registration network model is a UNet-like network, non-rigid registration is carried out on feature maps of different scales by using a spatial transformation layer, the registration feature maps and feature maps of adjacent scales in a decoder are fused, and registration fields of the adjacent scales are also fused, so that the registration capability of the model on large-deformation printed matter images is comprehensively improved. According to the method, the problems existing in image registration of part of industrial paper printed matters at present can be solved, and the registration effect is good.

Description

technical field [0001] The invention belongs to the field of defect detection of industrial printed matter, and in particular relates to a method and device for image registration of industrial printed matter based on deep learning. Background technique [0002] Image registration is an important research direction in the field of computational vision. The main purpose is to match two or more images under different conditions (illuminance, shooting angle and shooting position, etc.) to achieve geometric alignment. At present, image registration can be divided into traditional registration methods and image registration methods based on deep learning. Traditional registration methods include traditional rigid registration methods and traditional non-rigid registration algorithms. Among them, traditional rigid registration algorithms, such as feature-based image registration algorithms, search grayscale, edge and artificially designed image parameters in the image parameter sp...

AI Technical Summary

Problems solved by technology

However, paper printed matter is a non-rigid material, and it is easy to produce various deformations under the action of external force. Even if it can be fixed by clamps, it is difficult to obtain the exact same shape as the template image. Traditional registration methods are difficult to effectively solve paper printed matter. The problem of non-rigid image registration cannot fundamentally reduce the misjudgment caused by defect detection

Images