Few-shot out-of-distribution detection based on contextual cues

Abstract

This invention discloses a few-shot out-of-distribution detection method based on background cues. A pre-trained multimodal model is used as a baseline model to construct a few-shot out-of-distribution detection model based on background cues. The multimodal model extracts text features from learnable background text cue vectors and class cue vectors containing learnable class text cue vectors, respectively. These features are then combined with global and local image features to obtain local background similarity and local class similarity. Local background similarity is refined based on local class similarity, and then the background region is extracted. The few-shot out-of-distribution detection model based on background cues is trained using a training dataset. The sample to be detected is input into the multimodal model, and an OOD detection score is calculated to complete the out-of-distribution detection. This invention introduces background cues and refines local image similarity, improving the accuracy of background information extraction and the robustness of out-of-distribution detection.

Images