Visual Recognition Of Steel Plate Label Based On Attention Mechanism And Heat Map

Due to the particularity of the manufacturing process and operating environment,iron and steel enterprises usually use the method of spraying or marking on the surface of products to achieve production process tracking and product quality control.The common labels identification method combining video surveillance and manual interpretation in China has shortcomings such as hidden dangers to personal safety,low efficiency,and high investment cost.Combined with the project requirements of a steel plate intelligent labeling system in a steel enterprise,this paper studies the online detection and identification method of steel plate production line based on machine vision,and introduces artificial intelligence technology to solve the character recognition problems including uneven lighting,complex background,and super-long,multi-line printing.The main work are as follows:(1)Design and build an online imaging system for steel plate labels.For the imaging requirements of moving steel plates,a method combining multiple industrial cameras and LED light sources is adopted.For the problems of arbitrary stop,reciprocating motion during the movement of the steel plate and large-area capture at different positions of the steel plate,combined with the sensor and on-site control circuit signals,the camera is jointly controlled by the software and hardware trigger mode to accurately capture the image of the steel plate.(2)Recognition of side-sprayed steel plate printing based on group-reversal attention mechanism.To solve the problems of small size of side-sprayed characters,complex background,uneven illumination,etc.,Efficient Net V2 is used as the backbone feature extraction network to enhance the extraction ability of complex high-level semantic features.Drawing on the idea of the group-reversal attention mechanism refines the feature information,multiple group-reversal attention(GRA)blocks are added in the feature extraction process to improve the sensitivity to hidden side-spray characters.Finally,the Bi LSTM structure is used to predict characters from the refined feature sequence,which is transcribed by the CTC algorithm.Experiments show that the improved model significantly improves the identification accuracy of side-sprayed printing.(3)Steel plate upper surface spray printing and steel seal recognition based on the heat map analysis.Traditional text positioning models are difficult to detect whole super-long,multi-line surface spray printing,and image sequence-based text recognition methods are not applicable for steel seal that are easy to lose points and have poor imaging quality.This paper proposes to use the YOLOv5 model to detect single characters of surface spray and stencil labels,and combine the affinity between adjacent characters and heat map to arrange the discrete character results in an orderly manner.In addition,this paper proposes an image enhancement method based on character-level annotation,which improves the efficiency of data set production and reduces labor and time costs.Based on the above algorithms,the intelligent recognition software for steel plate labels is designed to realize real-time recognition of three different steel plate labels,and the mean of the average recognition rate has reached more than 95%.