| With the in-depth development of my country’s intelligent manufacturing2025,the development of industrial manufacturing to digitization,networkin g,and intelligence has become an inevitable trend.Relying on the continuous innovation of computer technology,computer vision technology has also developed rapidly,making it play a great role in promoting industrial automation production.Character recognition is also an extremely important link in actual industrial production.Because it is difficult for the human eye to work efficiently for a long time,and in some harsh production environments such as high temperature,humidity,and darkness,it will make the detection work more difficult.Therefore,character recognition technology based on machine vision has been widely used in industrial automation production.Before deep learning technology is immature,machine vision technology mostly relies on traditional image processing technology,but the industrial production environment is complex,and the consideration of recognition accuracy,precision and cost in production is relatively strict.Therefore,in some complex environments,based on Traditional image processing algorithms cannot solve these problems well.In recent years,the rapid development of deep learning technology has provided solutions to these problems.However,due to the difficulty of data collection in industrial production,this paper conducts an indepth analysis and research on the character recognition algorithm in industrial scenes under the condition of small samples,and mainly completes the following research contents:(1)The development history of machine vision technology based on image processing and deep learning and its application in the field of character recognition in industrial scenarios are analyzed.Analysis of data augmentation techniques based on generative adversarial networks.Comparing the advantag es and limitations of different algorithms is used as the preliminary investigation of the research content of this paper.(2)This paper explores traditional character recognition methods,analyzes the principle of Tesseract character recognition algorithm based on traditional image processing technology,and uses this algorithm to realize character recognition on character data in industrial scenarios.On the basis of analyzing the limitations of traditional image processing,this paper studies a charact er recognition algorithm based on deep learning.Aiming at the problem that deep learning network relies on a large amount of data for training,while the training data is small in industrial scenarios,and the category and non-standardized character style data are incomplete,three data augmentation methods including pixel-level,geometric transformation and generative adversarial network are analyzed..For the problem that the pixel-level and geometric transformation data augmentation methods have weak improvement in character recognition,this paper analyzes the data augmentation method based on generative adversarial networks,and studies and improves a small sample character data augmentation strategy in industrial scenarios.It consists of two modules :character detection and synthetic data generation.(3)By analyzing the characteristics of the character data used in this paper without pre-and post-semantics,the current mainstream target detection network YOLOv5 is used as a character recognition network,and its development process and algorithm principle are studied.The character detection module in the character data amplification strategy studied in this paper adopts the same target detection network as the character recognition module,and obtains a single character area and an ROI containing all characters through the network,and the obtained results are used for the subsequent synthetic data generation module.(4)The synthetic data generation module consists of a character style transfer module,a character erasure module and a character fusion module.The character style transfer module and the character erasure module are implemented using an image-to-image translation model based on generative adversarial networks.By establishing the character style transfer module training set,a large number of single characters with industrial scene style are generated.The character erasing module erases the characters in a large number of unlabeled original image character ROIs and retains the original features of the non-character regions to obtain a character-free background.In this paper,the cycle-consistent adversarial network is improved for the character erasure module,and an attention mechanism is added to the cycle-consistent confrontation network to optimize the effect of the network on character erasure.The character fusion module integrates a large number of generated single characters with industrial scene style into a character-free background to obtain a composite picture to realize data amplification.(5)Through the analysis of the character recognition algorithm based on deep learning,this paper uses the small sample character data augmentation strategy studied to analyze the surface of the rubber model of the braces,the metal surface of the charger,the plastic surface of the SIM card,and the SIM card,each containing 15 original pictures.Four groups of small-sample character datasets on the metal surface were augmented to generate 3,000 composite images.The recognition accuracy of the four groups of data augmented datasets trained on the YOLOv5 network was higher than that without data augmentation.10%-20% improvement.For the missing categories in the dataset,after processing the data augmentation strategy studied in this paper,only the synthesized characters are trained,and the single-category recognition accuracy AP can reach more than 90.(6)The research done in this paper is summarized,its limitations are analyzed,and the follow-up research that can be expanded is summarized and prospected. |
