Research And Application Of Object Detection And Instance Segmentation Technology In Industrial Vision

Industrial vision technology is a key technology that needs to be vigorously developed as planned in Industry 4.0 and Made in China 2025.Warehouse management is one of the important fields of industrial vision technology application.The most important aspect of warehouse management is the counting of goods entering and exiting the warehouse.This paper mainly focuses on the counting problem of a group of boxes with the same size that are packaged into a cargo pile and transported in and out of the warehouse by vehicles.At present,most enterprises still use manual labor to record the quantity of incoming and outgoing warehouses,which has the drawbacks of high cost,insufficient accuracy,and low efficiency.With the rapid development of artificial intelligence technology,intelligent counting algorithms based on computer vision technology have emerged.Compared to traditional manual management,intelligent counting algorithms have a series of advantages such as fast speed,high accuracy,long traceability time,fast and convenient queries,which can effectively improve the efficiency of enterprise cargo entry and exit management processes.However,there are still significant problems with the current computer vision technology in solving the problem of intelligent counting.Firstly,due to the complex background,it is difficult to accurately identify the precise position of the cargo pile to be counted in the field of view;The second reason is that there are many changes in the packaging methods of goods,and the gaps between the boxes are not obvious,making it difficult to identify individual boxes.In response to the above issues,researching a set of efficient and reliable intelligent counting key technologies has significant theoretical and practical value for the digitization and intelligence of logistics warehousing management.At present,existing object detection models,such as FasterR-CNN,YOLOv5,etc.,have found through experiments that in the application scenario of this article,there may be partial missing recognition of the edge area of the cargo pile,making the recognition results unable to fully include the target cargo pile.In addition,there may be misidentification of the background.For existing instance segmentation models,such as Mask R-CNN,SOLOV2,etc.Through experiments,it was found that the model cannot adapt to the enterprise warehouse environment with significant interference.Due to the deformation and tight arrangement of the boxes in the environment,the model’s cargo pile segmentation effect in this actual production environment is poor and cannot achieve the required accuracy.Given the above reasons.this article proposes a new machine vision cargo counting algorithm for application scenarios,which combines object detection YOLOv5 and semantic segmentation HRNet seg deep learning model to achieve accurate positioning of the input image cargo pile and accurate instance segmentation of the box.The new algorithm has significantly improved the accuracy and efficiency of cargo pile counting recognition,and has developed a counting platform based on the new algorithm,which has been successfully applied in the on-site production environment of enterprise warehouses.The main work and contributions of this article include:(1)Optimize YOLOv5 target detection algorithm,change the network loss function by modifying the data enhancement proportion method,so that the target detection network can quickly and accurately locate the position of the cargo pile in the current input image.(2)Propose a new instance segmentation algorithm,combined with the HRNet seg semantic segmentation network,to accurately and quickly obtain the area occupied by each box in the current cargo stack for different placement forms.On this basis,a precise and reliable box counting method was designed and implemented by combining the graphic region filling algorithm.(3)Based on intelligent counting algorithms,a set of intelligent counting platforms has been developed and deployed on site in enterprises,achieving good results.In summary,based on the application background,this article proposes a new machine vision cargo counting algorithm based on object detection and instance segmentation.It builds an intelligent warehouse counting platform that can be applied to actual production environments and is deployed in a ceramic enterprise warehouse management application in Guangdong.The trial operation has achieved satisfactory results.Finally,this article provides a prospect for the existing problems and future directions of the algorithm.