Delver Hamed : Content Based Image Retrieval Using BEIT Transformer and Detic

Abstract

I present DHam, a new and exact unsupervised learning model for Content Based Image Retrieval (CBIR) that does not need any training data set. DHam is accurate especially when you deal with the multiple objects image with background (MOIB). This is the first time that pre-trained Detic and pre-trained of a self-supervised based image transformer (SSIT) BEIT, have been mixed for CBIR. First, I use pre-trained Detic to detect image objects. Then I extract every object's feature with pre-trained BEIT. DHam shows its superiority when search is accomplished amongst multiple objects images with background (MOIBs). Besides, DHam Test results are compared with pure BEIT and pure ResNet CBIR models. On the other hand, it is not a fast model. It takes around 19 and 273 seconds to compare the input image with 44,891 and 1,868,672 features respectively. Compared to state of the art CBIR systems, DHam may bring irrelevant images but is less likely to miss the target similar image.