Research On Image Out-of-distribution Detection Method Based On Deep Learning

Out-of-distribution detection is an important task in image processing,and its basic goal is to detect test data with different class labels from the training data,which are significantly different from the training data in terms of distribution.In real scenarios,there are some unknown data that are different from the training data due to noise,interference,etc.These unknown data may negatively affect the performance of the model.By performing out-of-distribution detection,it can help the model identify the unknown data more accurately and make correct predictions.Image out-of-distribution detection has a wide range of applications in many fields,such as autonomous driving,medical image analysis,etc.In these fields,accurate recognition and processing of unknown data by the model is crucial,so image out-of-distribution detection is gradually becoming one of the important directions of deep learning research.Since its emergence,the problem of out-of-distribution image detection has received extensive attention from researchers in related fields,and detection methods based on classification,density,distance,etc.have been proposed.However,due to the influence of out-of-distribution data training dependency,out-of-distribution detection single-layer representation dependency and out-of-distribution detection diversity,most detection methods have unstable out-of-distribution detection performance and poor generalization ability for unknown out-of-distribution data when performing tasks in different out-of-distribution detection scenarios.In this paper,by analyzing the above problems,we propose two improved out-of-distribution detection models,and the main contributions are summarized as follows:(1)An out-of-distribution detection model based on subspace reconstruction is proposed.The model uses only in-distribution data for training in the training phase without deliberate division of out-of-distribution data,which solves the problem of out-of-distribution data training dependency.And the model introduces the out-of-distribution detector into the middle layer of the network to solve the single-layer representation dependence problem of out-of-distribution detection.The sample features are reconstructed by the hierarchical feature detection module of the model,and the feature reconstruction difference value between the original features and the reconstructed features after projection is used as the uncertainty score of the out-of-distribution detection.Since the model is trained using only in-distribution data,resulting in out-of-distribution data being difficult to be reconstructed during testing,the feature reconstruction difference value of in-distribution data is low,while the feature reconstruction difference value of out-of-distribution data not seen by the model is large,thus achieving effective identification of out-of-distribution data.Experiments on four out-of-distribution datasets,including SVHN,show that the model has better out-of-distribution detection performance compared with other detection methods.(2)An out-of-distribution detection model based on multi-level feature detection is proposed.The model adds a grouping softmax detection module to the subspace reconstruction out-of-distribution detection model to reduce the decision boundary between indistribution data and out-of-distribution data by grouping,aiming to ensure the correct classification of in-distribution data while enhancing the model adaptability.The model introduces other classes in each semantic group,and then calculates the probability distribution within each semantic group by grouping softmax,and selects the smallest other class score in each group as the final out-of-distribution probability score of the image.Comparative experiments are conducted on four out-of-distribution datasets such as LSUN,and the experimental results show that the out-of-distribution detection performance and model adaptation of the model have been further improved.(3)Examine the performance of the model under the diversity scenario of out-ofdistribution detection.An intuitive interpretation of the detection results for different outof-distribution data is firstly achieved by showing the detection of the hierarchical feature detection module.Secondly,experiments are conducted with increasing number of samples in the in-distribution training set,increasing number of categories,different backbone networks,different grouping strategies,and different module compositions as conditions,respectively,and the experimental results show that the multi-level feature detection models are stable under different out-of-distribution detection conditions.Finally,the two out-ofdistribution detection models are summarized for the applicable scenarios.