Image Data Annotation And Recognition Based On Weakly Supervised Deep Learning

With the in-depth development of artificial intelligence technology,the field of computer vision has made a lot of achievements,especially in the field of computer vision recognition,excellent image target recognition models emerge in endlessly,but there are still many challenges in the target recognition task.On one hand,the training of computer vision recognition model often needs the support of a large number of image data,on the other hand,accurate image instance level annotation needs extremely expensive human cost.In order to provide better basic support for the target recognition of computer vision and automatic driving perception field,this paper designs a labeling software for traffic environment,and manually annotates the collected image data in real traffic environment.Based on this data set,this paper designs a target recognition model in road traffic environment.Furthermore,in view of the challenges existing in the real road traffic environment,this paper optimizes the designed model,which improves the performance of the model significantly.The costs of accurate image annotation is huge,and the high cost is not conducive to the acquisition and application of computer vision model.Facing this challenge,this paper designs a method to reduce the workload of image data annotation.This method combines weakly supervised learning with active learning,initializes the model with weakly supervised data,selects the most effective samples to improve the model by active learning,and trains the model iteratively,which effectively reduces the amount of image annotation and the costs of image annotation.Similarly,focuses on how to reduce the labeling workload of image data,this paper also improves the weakly supervised target recognition model,designed a weakly supervised recognition model integrating attention mechanism,which makes the model pay more attention to key channel and spatial features,enhances the expression ability of the model features and improves performance of the weakly supervised model.