Research on chemical substance identification methods driven by miniature spectral imaging

Abstract

Miniature spectral imaging is convenient for in-situ chemical identification, but most portable devices have few spectral bands, unstable illumination and strong sample-shape effects. A concise recognition method is presented in this paper that integrates reflectance correction, optimal band selection, spectral-spatial feature construction and a light-weight attention fusion module. The image cube is corrected for dark-field/white-reference calibration and band registration. Spectra are then normalised, smoothed and denoised, and redundant bands are removed before classification. Fuse local spectral and spatial features to obtain class labels and confidence scores. Based on the experiment results for powders, crystals and thin films, the overall accuracy of the method is 96.1%, the macro-F1 score is 0.957, and the per-frame latency is 18.7ms. According to the above data, the new pipeline can be used to construct a relatively low-cost mobile chemical detection device that supports calibration and quality-control functions.