Pixel-driven mobile phone operation interface text detection method

Abstract

The invention discloses a pixel-driven mobile phone operation interface text detection method. The method comprises the following steps: 1) generating a mobile phone operation interface image; 2) improving a text detection model; and 3) correcting a detection box error based on a projection method. According to the pixel-driven mobile phone operation interface text detection method, the problem oflack of text marking data of an existing mobile phone operation interface is solved; the position and font size of the generated data are random within a certain range, and the generalization of themodel can be improved by training; an up-sampling layer is redesigned for the output of ResNet50, so that the deep feature proportion is improved; aiming at the condition that the mobile phone operation interface scene hardly has angle deviation, an angle regression branch of a regression part of the detection frame is removed and a loss function is correspondingly modified to improve the networkconvergence speed; the detection frame error correction method based on the projection method can assist in correcting the detection frame deviation under a pure color background, and the problems ofincomplete character detection, large blank area or the like are reduced as much as possible.

Description

technical field [0001] The invention relates to the technical field of image processing. Background technique [0002] The mobile phone interface contains three elements: text, icons, and pictures. In order for the computer to understand the semantics of all elements in the interface from bottom to top, it is necessary to divide the interface into several small pieces with independent semantics. Because the texture of the text area is complex and the aspect ratio is uncertain, it is usually impossible to completely identify the text area using a general-purpose target detection method, so a specific text detection method is required for text segmentation. [0003] At present, text detection methods are divided into methods based on the underlying features of images and methods based on deep learning. Methods based on the characteristics of the image itself (such as MSER, ECCV2008 and SWT, CVPR2010) detect text by analyzing the distribution of spot areas and strokes in the i...

AI Technical Summary

Problems solved by technology

The methods proposed by researchers recently mainly focus on the distortion and projective transformation of text in natural scenes, but the text in the mobile phone interface almost does not have this problem, and it is not suitable for the application scenario of this scheme.

Images