ISRToken: Learning similarities tokens for precise infrared spectrum recognition model via transformer

Abstract

Infrared spectrum recognition (ISR) has been widely used in the fields of biomedical imaging and cancer diagnosis. However, existing methods cannot deal with extreme infrared spectral random noise and band overlap. To address these challenges, we identify three cues from infrared spectrum, namely, neighborhood similarities, significant spectral structure changes, and key minority relationships. To leverage the observed findings, we propose a novel key minority relationship-aware method based on the Transformer architecture in which the spectral section relationships can be learned. Specifically, we design several orientation tokens to explicitly encode the basic similarities regions. Meanwhile, a novel token guide multi-loss function is designed to guide the similarities tokens as they learn the desired regional similarities and relationships. We evaluate the proposed method on three challenging benchmark ISR datasets. Experiments show that our method achieves better performance compared with state-of-the-art methods.