Research On Character Recognition In Industrial Environment Based On Deep Learning

Optical Character Recognition(OCR)technology began in the mid-1960 s.After the emergence of deep neural networks,the recognition object developed from printed characters to natural scene characters.At present,OCR based on deep learning has become an important research topic in the field of computer vision.With the introduction of Made in China 2025,which promotes the development of Chinese industry towards informationization,the application of character recognition technology in the industrial environment has received extensive attention.Different from high-resolution and high-definition document character images,the characters in complex industrial environments often have low-quality conditions such as noise,scratches,and blurs,which brings greater challenges to character recognition and severely restricts industrial information process.This paper divides industrial environment characters into two types: production line collection character and long-distance imaging collection character.The difficulty of recognition is solved by constructing different deep neural network models.The main tasks are as follows:(1)In the case of noise,scratches and other interference information in the characters collected by the industrial production line,the CRNN character recognition model based on the channel attention mechanism is constructed through the method of squeeze-stimulus-recalibration.The channel attention mechanism increases the weight of key feature channels to improve the expression ability of character features in the feature map,thereby improving the robustness of the model.The experimental results show that the CRNN model based on the channel attention mechanism significantly improves the recognition accuracy of industrial production line characters.(2)For the low resolution and poor definition of character images caused by longdistance imaging in an industrial environment,Plugnet,a character recognition model fused with a super-resolution module,is used and improved.By using Leaky Re LU to enhance the information transfer ability in the feature extraction process;by using the multi-task loss function of L1 and MS-SSIM to improve the model's correction method for the convolutional layer parameters in the backpropagation process,the mapping relationship of the model is optimized.Using natural scene characters to replace longdistance imaging characters in industrial environments,the improved model has improved recognition accuracy for seven data sets such as SVT.