Research And Implementation Of Natural Scene Text Detection Algorithm Based On Lightweight Network

In today's big data era,effectively acquiring and using textual information from natural scenes is of great significance to information retrieval and intelligent life.At present,deep learning technology is developing rapidly,and has been widely used in some image fields such as target detection and face recognition.These methods use deep convolutional networks to extract image features and use these features for detection and classification tasks.At present,the mainstream text detection methods for some natural scenes are improved by classic target detection methods,but most of the target detection algorithms are based on Anchor.In order to improve the recall rate,these methods will generate a large number of Anchors and bring a large number of calculations,which will bring great losses in real-time and memory consumption.Not only that,the design of parameters such as the side length and aspect ratio of Anchor also requires a lot of experiments to verify the validity of the parameters.To solve these problems,an Anchor-free natural scene text detection algorithm is designed.The algorithm removes the original Anchor,establishes pixel-level classification and regression tasks,and returns the upper and lower values of the distance border on the pixels of the text box.In terms of feature extraction,the algorithm uses a lightweight network Shuffle Net V2,which has a deeper depth and a larger receptive field.At the same time,the OHEM algorithm and shrink operation are used to solve the imbalance problem of positive and negative samples and the boundary problem.The algorithm is not only faster than the original algorithm,but also greatly improved in accuracy.On some public data sets,the algorithm is 6-7 percentage points higher than the previous method,which verifies the effectiveness of the algorithm.Finally,a natural scene text detection system is designed and implemented,the purpose of which is to monitor in real time whether the live content of outdoor anchors contains sensitive information,and display the detection results on the client to remind the administrator to operate.The evaluation standards and test results obtained by ICDAR show that the system can monitor live content well,which indicates that the designed algorithm has high value for use.