Research On Out-of-distribution Detection Algorithm Of Vision Based On Deep Learning

With the development of deep learning,artificial intelligence has been applied in more and more fields.However,many algorithms still have great security problems in practical application and deployment.Existing supervised algorithms are generally based on the closed-world assumption that training and test data originate from the same distribution,and when algorithms encounter deviations from the training data distribution,they often produce incorrect and overconfident predictions.The purpose of out-of-distribution detection algorithms is to catch erroneous predictions ahead of time by analyzing the data distribution and detecting potentially erroneous samples.Therefore,out-of-distribution detection algorithms have important research significance and research value in safe and robust deployment in various fields.In this paper,we mainly study the visual out-of-distribution detection method based on deep learning,which is oriented to the two fields of video and medical images.According to the different characteristics of video and medical images and different task requirements,different out-of-distribution detection methods are designed.Experimental results show that the proposed method improves out-of-distribution detection in video and medical images.The main contents and innovations of this paper include:· The applicability of image out-of-distribution detection algorithm in video is studied,and a video out-of-distribution detection method based on supervised contrastive learning is proposed.Supervised contrastive learning can increase the distance between heterogeneous samples and reduce the distance between similar samples,making it easier to detect out-of-distribution samples.Experiments show that supervised contrastive learning makes the intra-class features more compact and improves the effect of detecting samples outside the distribution.· A reconstruction-based method for out-of-distribution detection in medical images is studied,and a foreground mask reconstruction-based method is proposed to learn high-level context-based semantic information.Aiming at the characteristics that medical images are sensitive to details such as texture,an out-of-distribution detection method based on texture reconstruction is proposed.Through ensemble learning,the proposed algorithm achieved fourth and fifth place on the sample-level and voxel-level tasks of MICCAI’s MOOD 2020 challenge,respectively,and first place on the sample-level task on the abdominal CT dataset name achievement.