A lightweight spectral qualitative analysis method based on Res-Inception.

Abstract

Infrared spectroscopy is widely used for qualitative classification tasks in pharmaceutical, food, agricultural product and other fields, owing to its inherent technical advantages of rapid detection, non-destructiveness, reagent-free and pollution-free operation, as well as reliable in-situ detection capability. However, traditional qualitative spectral analysis methods rely heavily on manual feature extraction and shallow machine learning models, making it difficult to fully capture the nonlinear features embedded in high-dimensional spectral data, with their generalization performance severely constrained by prior domain knowledge and dataset characteristics. Although existing deep learning methods have improved classification accuracy, most models are custom-developed for specific tasks in a single domain, resulting in insufficient multi-domain generalizability, while their large parameter count and computational complexity restrict edge deployment on portable spectral instruments and engineering-scale promotion. To address the above limitations, this paper proposes DSRINet-1D, a lightweight one-dimensional (1D) spectral classification network integrating depthwise separable convolution, Inception multi-scale structure, and nonlinear residual connection, which drastically reduces model computational complexity and parameter count while ensuring classification accuracy and universality. A Feature Screening and Weight Assignment (FSWA) attention mechanism is further designed, which dynamically optimizes the weight allocation of multi-scale features, significantly enhancing the model's feature representation capability with only a slight parameter increase, thus achieving an efficient balance between classification performance and lightweight design. Experimental results on three cross-domain public datasets, namely Tablets, fruit juice, and meat, demonstrate that the FSWA-embedded DSRINet-1D achieves classification accuracies of 99.355%, 98.474% and 99.726% respectively, with comprehensive performance significantly superior to traditional machine learning methods and mainstream deep learning models. Compared with the baseline Res-Inception model, the basic DSRINet-1D reduces parameter count by 42.4% and floating-point operations (FLOPs) by 94.4%, exhibiting a remarkable lightweight advantage. This model requires no complex manual feature selection or preprocessing procedures, boasts strong generalization capability, and is easy to deploy for rapid field applications. It also provides a novel research insight for spectral classification tasks with different input forms, and explores a two-dimensional spectral modeling method based on dual-band transformation.