EAST algorithm optimization method

Abstract

The invention belongs to the technical field of natural scene text detection algorithms, and particularly relates to an EAST algorithm optimization method, which comprises the steps of 1, carrying out image preprocessing; step 2, inputting into a feature extraction network for feature extraction; step 3, carrying out feature fusion; 4, inputting the data into a BLSTM network; and step 5, outputting a result. By optimizing a mixed feature pyramid structure, four layers from top to bottom are combined and improved to only combine the first two layers, and then the combined layers are fused with a path from top to bottom. A feature extraction and feature combination structure of an original EAST algorithm are replaced with the optimized mixed feature pyramid structure. The method solves the problems that the EAST algorithm does not pay enough attention to the small receptive field text, the small receptive field feature map processing is less, and the subsequent processing is not facilitated. The prediction error of the EAST algorithm is greatly reduced, and the prediction precision of the EAST algorithm is improved.

Description

technical field [0001] The invention belongs to the technical field of natural scene text detection algorithms, in particular to an EAST algorithm optimization method. Background technique [0002] Traditional text detection methods and some text detection methods based on deep learning are mostly multi-stage structures (multi-stage), and each level needs to be tuned during training, which will inevitably affect the final positioning effect, and it is very consuming Time. In response to the above problems, Questyle Technology proposed the EAST algorithm (EAST: An Efficient and Accurate Scene Text Detector) in 2017. The EAST algorithm proposed an end-to-end text detection method, eliminating multiple stages in the middle (such as candidate region aggregation, text word segmentation) , post-processing, etc.), directly predicting text lines. The structure of the EAST model is simple, and the focus is on the use of FCN full convolutional neural network structure and loss functi...

AI Technical Summary

Problems solved by technology

Since the feature maps of feature merging are all derived from the feature maps of the feature extraction network, the two feature extraction paths are exactly the same (such as image 3 As shown, the feature extraction is actually only performed once) only in the direction of the feature merging part is different, so it will only bring a small increase in the amount of calculation in the feature merging part, which not only solves the problem of insufficient attention to the text in the small receptive field of the EAST algorithm, and the small feeling Less processing of wild feature maps leads to insufficient abstraction of features, which is not conducive to subsequent processing, and will not bring too much burden to the calculation of the network

Images