Data enhancement method applied to signboard photo identification

Abstract

The invention provides a data enhancement method applied to signboard photo recognition, and the method comprises the steps: S1, carrying out the signboard image recognition of an image uploaded by a user, and filtering out non-signboard images; s2, constructing a training sample composed of a signboard photo and a random mask to carry out U-net network model training, and obtaining a repair model; s3, recognizing a light spot in the signboard image by adopting an image binarization method, and repairing the light spot; s4, identifying a signboard area P1 in the signboard image P and a character area in the signboard area P1; s5, determining an effective character area; s6, segmenting a character foreground style in the signboard area P1, extracting characters of random fonts from the dictionary to form an image Font, and replacing the characters in the effective character area with the image Font; and S7, fusing with the signboard area P1, and replacing the original signboard area in the signboard image. The problem that the sample size in model training is insufficient is solved.

Description

technical field [0001] The present application belongs to the technical field of data processing, and particularly relates to a data enhancement method applied to signboard photo identification. Background technique [0002] With the continuous popularity of artificial intelligence algorithms in the field of image recognition. Existing artificial intelligence technology, such as deep neural network and other algorithms, requires a large number of samples to train the algorithm model. Due to the particularity of signboard photos, all image samples need to be taken by users themselves, and the quality of data obtained by the network is also uneven, so the number of samples is usually insufficient. Insufficient sample size will challenge signboard recognition accuracy. [0003] At the same time, the signboard is designed according to the user's preferences, so different signboard photos involve different fonts, different layout methods and different design styles. The diversi...

AI Technical Summary

Problems solved by technology

Insufficient sample size will challenge the accuracy of signboard recognition

[0003] At the same time, signboards are designed according to user preferences, so different signboard photos involve different fonts, different layout methods, and different design styles. The diversity of these data will also result in insufficient samples under this style

[0004] Existing data enhancement techniques including tilting, zooming, flipping, etc. increase the sample size of the data, but affect the final image recognition effect; fusion text data synthesis can solve the problem of insufficient data in the text recognition stage, but cannot solve the problem of text recognition. The problem of insufficient data in the detection phase

Images