Design And Implementation Of Fast Labeling Software For Road Waste Image Datasets

Machine vision technology is one of the research hotspots in the field of optical engineering,and deep learning is an important support point of machine vision technology.Deep learning training often needs to annotate a large number of image data sets,and its production needs a lot of manpower and material resources,which is a boring and tedious work.In this paper,according to the actual requirements of scientific research projects and the image characteristics of road garbage under different lighting conditions,a set of fast labeling software system for road garbage image data set is developed.In this paper,two fast labeling methods are designed and implemented in the fast labeling software system of surface garbage image dataset.The interactive fast labeling function is realized by using the mixed Gaussian background modeling algorithm and the differential color information background modeling algorithm.The algorithm can automatically select and mark all foreground objects in the current area by simply selecting the area of interest,and the user can adjust the category and size of the boundary box twice.In addition this paper also designs a further deep learning AI collaborative tagging function,user manual annotation some pictures first,then this batch of pictures for network training,and training after the network model is used to automatically for subsequent image annotation,then the user can adjust and modify,this way can effectively improve the labeling efficiency of mass image data.The fast marking software developed according to the above ideas has been used to mark the road waste data set of the intelligent sweeper road waste identification system.Compared with the mainstream manual labeling software,the actual test results showed that the labeling speed was more than twice as fast as that of the mainstream manual labeling software.After the training of 150 images in the actual road garbage data set,the MAP(average accuracy)of the AI collaborative labeling reached more than 0.65,and the average time for the image with5 million pixels was 76.59 ms.In the labeling test of pavement waste data set,the average accuracy of fast labeling based on the mixed Gaussian background modeling is 0.896,and the average accuracy of fast labeling based on the differential color information background modeling is 0.756,both of which can effectively improve the labeling efficiency.