Research On Container Code Recognition Technology Based On Computer Vision

The existing container code recognition system can only meet the needs of container code detection and recognition in ordinary scenes,and has low recognition accuracy and poor antiinterference ability in complex scenes such as strong light irradiation,container oil pollution,other character interference,and container code skew.In order to improve the accuracy of container code recognition in complex scenes,this paper conducts in-depth research on the existing text detection and recognition technology and container code recognition technology,summarizes the advantages and disadvantages of these technologies,and proposes a container code recognition system with high recognition accuracy in any scene.The main work is as follows.(1)The design of the code area detection module.In order to improve the accuracy of code area detection,the detection module in this paper uses both the traditional algorithm based on improved MSER and the deep learning algorithm based on EAST to achieve code area detection,and a synthetic module is designed to combine and optimize the detection results of the above two methods to obtain the final detection result.In the MSER detection module,the edge is first enhanced by image sharpening,and MSER detection is realized based on edge gradient enhancement to improve the detection effect.For the candidate regions detected by MSER,most of the interference regions are filtered out by improved NMS algorithm and multi-mechanism filtering strategy.In the EAST detection module,the Advanced EAST detection model is used to solve the problem that EAST has poor prediction effect on long texts.In the comprehensive module,the detection result of Advanced EAST is used to filter out all interference regions of MSER,and then the detection results of the two methods are combined and optimized.The synthetic module makes full use of the accuracy of MSER and the robustness of Advanced EAST to improve the accuracy of the overall code area detection,avoids the situation of missing characters and the situation where the detection box border overlaps with the characters.(2)The design of the code recognition module.In order to improve the accuracy of code recognition,this paper uses both the Le Net-5 recognition module based on character segmentation and the end-to-end CRNN recognition module to realize code recognition,and a synthetic module is designed to combine and optimize the recognition results of the above two methods to obtain the final recognition result.In the Le Net-5 recognition module,the Radon transform and the improved rotating projection method are used firstly to correct the tilt of the box code image,which improves the recognition effect of the module on tilted images.Then,the improved projection method is used to realize the segmentation of box code characters,which has a good segmentation effect when the box code characters are stuck or broken.Finally,the Le Net-5 network structure is improved to make it suitable for the container code recognition scene in this paper.In the end-to-end CRNN recognition module,the STN network is added to solve the image tilt problem,and the CRNN convolution layer is changed from VGG16 to Dense Net to extract richer character features,and the LSTM is changed to a Bidirectional LSTM for to fully utilize text context information,improve the recognition effect and enhance the robustness.The synthetic module is realized based on the box code design rule and the verification rule,which can obtain more accurate recognition results.The experimental test and data results show that the MSER method alone has high accuracy but low recall,and the Advanced EAST model alone has high recall but low accuracy.The synthetic detection module in this paper can achieve both high robustness and high accuracy,the recall rate reaches 93.23%,and the accuracy rate reaches 89.36%,which proves the effectiveness of the synthetic detection module.In addition,the recognition modules based on Le Net-5 and CRNN have only 87.72% and 83.25% recognition accuracy respectively,while the recognition synthetic module designed in this paper improves the accuracy of box code recognition to 94.17%,which proves the effectiveness of the recognition comprehensive module.