Scene Image Text Detection Based On Deep Learning Method

With the arrival of the information age,the number of images and videos is explosively increasing.Extracting information from massive data has become the focus of researchers in the field of artificial intelligence.Text information in scene images plays an important role in image retrieval,intelligent navigation and other applications.Therefore,the research on text detection and recognition algorithms for scene images is of great value.Great progress has been made in recent years,but it is difficult to meet the actual needs in terms of detection accuracy and speed.In practical applications such as Internet image text retrieval,the proportion of scene images containing text is small.To achieve fast text detection for scene images,this paper presents a scene text detection algorithm that fuses text/non-text scene image classification.First,classify and judge whether the scene image contains text,and then only perform text detection on the image containing text,so as to improve the performance of scene image text detection.The main work of this paper is as follows:(1)Based on the analysis of current deep learning image classification algorithms,a Light VGG network is designed to classify text/non-text scene images.In addition,based on the idea of grouping convolution and multi-scale feature extraction,a scaleadaptive module is designed,which combines with Light VGG to ensure high speed and low parameter quantity,and to achieve high accuracy in text/non-text scene image classification.(2)To further improve the classification performance,by introducing the activition boundary knowledge distillation,the classification accuracy is greatly improved while guaranteeing the high speed and low parameters of the network.In addition,the USlim module is introduced to solve the problem that network capacity needs to be adjusted flexibly in network deployment,which enables a single network to adjust capacity flexibly and improves network efficiency.(3)A scene image text detection algorithm that fuses text/non-text image classification is implemented.High-precision scene image text detection is achieved by using efficient feature extraction networks and loss functions.The detection algorithm fuses effectively with the text/non-text scene image classification module,which improves the speed of scene text detection greatly while ensuring the accuracy of detection.The algorithm is tested on the public databases Text Dis and ICDAR2015.The experimental results show that the scene text detection algorithm after fusing the text/nontext image classification algorithm improves the inference speed on GPU and CPU by more than two times.This fully demonstrates the effectiveness of the algorithm.