| Rechargeable lithium battery has become the most widely used hardware power storage unit with its excellent performance and reliable safety,such as small size,light weight,high specific energy,etc.With the rapid popularization and iteration of smart phones,there is a strong demand for the production and processing of mobile phone lithium batteries.At the same time,as the production process continues to improve,higher requirements are put forward for the quality of lithium batteries.However,in assembly line processing,it is difficult to avoid damage to the appearance of the battery due to defects in the production process,collisions or extrusion of objects.The current solutions of enterprises are mostly to set up specialized workers to sort defects according to different degrees and categories.This method has high labor costs,and the result of defect classification is subjectively affected,which is not conducive to obtaining valuable conclusions based on defect feedback to feed back the improvement of the production process.In response to the requirements of appearance defect detection in mobile phone lithium battery industrial assembly line production scenarios,this paper designs a complete set of practical solutions including image acquisition system and defect detection and classification algorithms based on computer vision.Images are collected from four views of front view,squint,prism,and tabs with five cameras,and use label Image to calibrate the battery location,defect type and location data on nearly26,000 images.Defect detection for eleven types of defects such as ear breaks whitishness,heterochromatic contamination,leakage crystallization,deep pits,liquid gas swelling,body deformation,surface scratches,skewed tabs,super high tabs,super high S and etc.The overall algorithm solution is mainly divided into the following three parts:First,a battery positioning algorithm based on the Faster-RCNN+FPN model framework is proposed,using nearly 20,000 manually annotated pictures as a data set to identify the specific location of the battery in the image sample,and remove most of the background content,and eliminate The adverse effect of background objects on defect detection.Second,the defect detection algorithm model is constructed on the basis of the Faster-RCNN+FPN framework,and the target detection task is completed in two stages,namely,"candidate frame acquisition" and "detection target classification".Approximately 20,000 fully manually calibrated mobile phone lithium The battery pictures are used as a data set,and the image samples cover non-defective qualified batteries and unqualified products with nine types of defects.The original picture resolution is 3840pix×2748pix.Third,the traditional image processing algorithm based on Open CV is used to detect the three kinds of ear-type defects with poor model detection results,and the detection accuracy is greatly improved.With the experiment of nearly 800 test samples,the recognition accuracy rate of the battery positioning model is about 97.3%,the average accuracy rate of the 6 types of defects in the pool body is 86.1%,and the average accuracy rate of the 3 types of ear-type defects is 84.5%.The calculation time for a single sample of the battery positioning model is about 200 milliseconds,the calculation time of a single sample of the defect detection model is about 400 milliseconds,and the calculation time of the defect detection auxiliary algorithm is 50 milliseconds for a single sample.The calculation time and calculation accuracy rate have reached the industrial level.The standard of pipeline application in the scene verifies the effectiveness and reliability of the solution. |
PDF Full Text Request