The invention discloses a multi-dimensional recognition method for out-of-distribution and misclassification detection, and belongs to the technical field of
computer vision and
deep learning. In view of two major core challenges of OOD (Out of Distribution) detection and Misclassification Detection of a
deep learning model in an open world scene, the invention provides a unified multi-dimensional recognition framework. According to the method, five Vision
Transformer models (BEiT, BEiTv2, DINOv2, MAE and ViT-Sup) based on different pre-training strategies are adopted to extract multi-view features in parallel, four kinds of complementary detection signals are constructed, the feature anomaly of a
signal 1 and the
mahalanobis distance of a
signal 2 form an out-of-distribution detection
signal group, and the prediction uncertainty of a signal 3 and the
decision boundary distance of a signal 4 form a misclassification detection signal group. Through an adaptive fusion strategy, confidence coefficient calibration is improved through temperature calibration, dimensional difference is eliminated through statistical normalization, fine utilization of signals is realized through dynamic weight calculation, and a comprehensive
score is generated. And performing three types of judgments based on the comprehensive
score: in-distribution normal samples, out-distribution samples and in-distribution misclassification samples. According to the method, joint modeling of the two types of problems is realized, the performance of distribution outside detection and misclassification detection tasks is remarkably improved, and a technical guarantee is provided for reliable deployment of a
deep learning model.