Research On Industrial Character Detection And Recognition Methods In Complex Backgrounds

With the rapid development of computer information technology in recent years,a huge amount of data will be generated,among which there will be a large amount of image data constantly generated,stored and transmitted.It is very important to analyze and process these image data to calculate the potential value.The use of deep learning to recognize characters in pictures is a very valuable and popular research direction,and nowadays character detection and recognition in natural scenes has made great research progress,while character recognition in industrial direction is still in its infancy.The main purpose of character recognition in the industrial direction is to automatically recognize the information of the workpiece number,so as to better record,track and detect the related workpiece products.The main work and research results of this thesis are as follows.(1)Character detection algorithm: To address the problem that the existing character detection algorithm is ineffective in detecting industrial characters with extreme aspect ratios,this thesis first produces a character detection dataset and then improves the PSENet network by applying a deformable convolution that can expand the sensory field to the FPN feature extraction network to solve the target problem.psenet is a new instance segmentation network that has two advantages.First,psenet,as a segmentation-based method,is able to localize text of arbitrary shape.Second,the model uses a progressive scale expansion algorithm which can successfully identify adjacent text instances.(2)Character recognition algorithm: To address the problem of poor industrial character recognition performance using existing algorithms,this thesis proposes a convolutional recursive neural network based on residual structure and "Squeeze-and-Excitation" block,named RS-CRNN.The feature extraction of this method is inspired by Res Net and SEnet,and RSblock is introduced in the feature extraction stage.GRU recurrent units are used for sequence label prediction.RS-CRNN can extract richer semantic information and thus achieve better character recognition prediction.(3)Character detection and recognition system: Based on the analysis and implementation of the above algorithm,we designed an Android software based on java language to make it our algorithm idea can be implemented on mobile.The main functions are as follows: photo function,recognition function,user management,database management,etc.The software runs smoothly,has a simple interface and good interactivity,which can better facilitate the management personnel to identify and manage related artifacts and greatly improve.