Scene text segmentation method based on weak supervision deep learning

Abstract

The invention provides a scene text segmentation method based on weak supervision deep learning, and the method comprises the following steps: enabling a scene picture to be superposed with any text to generate a scene text picture, generating a training sample, and labeling the training sample as the scene picture; performing feature extraction by using a convolutional neural network to obtain high-level semantics step by step; performing up-sampling through deconvolution to gradually recover the high-level semantic feature map to the size of the input image; carrying out multi-scale fusion on the feature maps output by the convolution layer and the deconvolution layer; activating the fused feature map to obtain a dichotomy black-and-white image of the scene and the text; setting a loss function for training; and corroding and expanding the scene text segmentation map obtained after training to generate a text region bounding box. The method does not need any strongly supervised pixel-level annotation sample, simply and efficiently solves the text segmentation problem in scene text detection, greatly reduces the algorithm cost, and improves the scene text segmentation efficiency.

Description

technical field [0001] The invention belongs to the field of image processing, and in particular relates to a scene text segmentation method based on weakly supervised deep learning. Background technique [0002] Extracting textual information from images of real-world scenes has become increasingly popular in recent years. Scene text detection (a technique for localizing text in natural scene images) plays an indispensable role in various text reading systems. Text detection in natural scene images is more complicated than general object detection. One of the main reasons is that the background of natural scene text images is extremely complex and diverse. For example, text can appear on flat, curved or wrinkled surfaces; near the text area There are complex interference textures, or non-text areas have text-like textures, etc. Because of the heavy noise, the algorithm has difficulty identifying text instances when locating bounding boxes. [0003] In the traditional met...

AI Technical Summary

Problems solved by technology

[0006] The technical problem to be solved by the present invention is to provide a scene text segmentation method based on weakly supervised deep learning, based on a fully convolutional semantic segmentation network, which does not require any strong supervised pixel-level labeled samples, and solves the problem in scene text detection simply and efficiently. The problem of text segmentation greatly reduces the cost of the algorithm and improves the efficiency of scene text detection

Images