The Study Of A Method For Recognizing Appearance Defects Of Can Based On 2D/3D Information Fusion

The beverage packaging market demand in our nation is huge.Aluminum cans have the advantages of good sealing performance and ease of use,and are widely used in beer and beverage packaging.Relevant national regulations and people’s consumption habits require that the bottom of the printed production date and batch number characters need to be correct to reflect the food filling line related information,the can body cannot have too obvious apparent defects,these requirements guarantee the health and safety of consumers and improve business interests.For the can filling line,on the one hand,the quality of laser coding equipment varies,the printing conditions are complex and easy to cause the bottom of the can character printing errors.On the other hand,due to the filling line processes are numerous,the instability of the transmission mechanism makes the production process of collision,friction,resulting in scratch,dent,deformation defects of the can body,which may lead to poor airtightness or abnormal pressure inside the can,and ultimately lead to food quality problems or production accidents.So,it is necessary to strictly detect and identify these defects before the can packaging food leaves the factory,so that companies can better control product quality.Machine vision detection,as a highly automated industrial detection method,is very suitable for continuous and rapid inspection of large quantities of cans,and also saves manpower and improves overall efficiency.This topic is oriented to two inspection stations of can bottom and can body in food packaging can filling line,to design a 2D/3D information fusion based can appearance defect recognition method,to promote the intelligent transformation of can filling line,to protect consumer health and corporate profits.The topic mainly focuses on the overall system design,can bottom character defect detection,depth information-assisted can body defect image enhancement,and can body defect recognition.The main research contents are as follows.(1)Overall design of can defect recognition system.For the requirements of multi-station and real-time of defect recognition,the topic works on the imaging scheme of can bottom and can body,build a distributed multi-source information acquisition system.Also design the overall process of duplex defect recognition algorithm and carry out the system software design based on external hardware trigger and multi-threaded parallel technology.The system software design is based on external hardware triggering and multi-threaded parallel technology.The defect recognition effect and system real-time performance are guaranteed.(2)Laser coding character defect detection method based on saliency semantic segmentation.Laser coding characters have low contrast,water stain noise,and weak character features,resulting in reduced character defect detection performance.To address this problem,firstly the Res18-UNet model with fused residual structure,improved feature chunking attention and multiple supervision mechanisms is proposed to enhance weak character features,improve image channel and spatial feature selection ability,and achieve saliency semantic segmentation of character targets,the F1 score reaches 0.9172 and can process 14.16 images per second.Secondly,a rotation correction and connected domain method is designed to extract individual characters for the varying character angles caused by random rotation of the cans on the production line.Then an ultra-lightweight character classification model is built to quickly classify each character with high accuracy to obtain the current character text,accuracy rate reached 99.557%.And compare it with the standard text to obtain the final detection results.(3)Depth information-assisted can body defect image enhancement method.The depth information of concave and convex defects such as dents and deformation is weakened due to the interference of complex backgrounds such as printed labels and logos on the can body,which can easily cause missed detection or mix with other non-convex defects.In order to improve the defect recognition performance,the design uses 3D depth information to assist in the enhancement and recognition of such defect images.Firstly,a multi-source data acquisition system is used to collect the lateral data of the can body in 3 directions;secondly,the calibration of the acquisition system is performed to unify the pose of the point cloud and to obtain the pose angle of the camera in the module.Then,in order to represent the lateral image information more intuitively,an improved cylindrical inverse projection model and a stitching method based on a priori information are built to expand and stitch the 2D map and the point cloud.For the problem that 2D/3D cameras have field of view differences,an alignment method based on field of view angle relationship is designed to realize 2D/3D data alignment,Meanwhile,the point cloud is downscaled and projected into a depth map,which is fused with 2D images to generate an enhanced map to complete the defect image enhancement.The results of multiimage recognition are fused to design a defect recognition method based on decision fusion.(4)Can body defect recognition method based on HPFST-YOLOv5.Aiming at the problem of low accuracy of defect recognition under the condition of serious noise interference such as printed label and LOGO in the 2D image of the can body,the traditional one-stage target detection model YOLOv5 is improved.A multi-level swin transformer multi-headed selfattention feature extraction structure is built as the backbone of the network to suppress clutter noise interference while strengthening the focus on the defect target.For the weakness of defect edge information caused by image motion blur and interpolation blur,a high-pass filtering guider is designed to emphasize the defect high-frequency edge information and guide the network backbone to optimize feature extraction,In the detection output part,a lightweight decoupled detection head is designed to make the network more focused on regression and classification outputs,respectively,to further improve the 2D image and enhanced map defect recognition accuracy.The m AP of HPFST-YOLOv5 for 2D image defect recognition is improved by 3.3% compared to YOLOv5 s,reaching 0.708.The final decision is fused with the detection results of depth map,and the can body defect recognition performance reaches the best,with m AP improved to 0.939.Based on the above can appearance defect recognition algorithm,the system software of this topic was developed using C++ combined with external libraries such as Open CV,QT and Tensor RT,containing user management,defect recognition automatic operation module,parameter teaching function module and data management module,which makes the algorithm of this topic has the solution of operation and deployment with friendly human-machine interaction.Based on the built laboratory collection system,a 6-week continuous process validation experiment was conducted to collate statistics,operating efficiency,and error analysis of defect detection results of duplex stations.The results show that the accuracy rate of character defect detection of the can bottom is higher than 97% counting by workpieces,the missing detection rate and over detection rate of the can body is less than 4.78% and 8.22%respectively counting by defects,and has good real-time performance.The recognition method of can surface defects based on 2D/3D information fusion in this topic can be better applied to the actual industrial inspection scenario of can duplex defect recognition task with high accuracy and speed.