Infrared Images Spectra Multi-class Classification Model Based on Deep Learning

Abstract

The classification of Fourier Transform Infrared spectra images is crucial in chemometrics. This paper proposes an efficient model based on deep learning approaches for enhancement and classification of the Fourier Transform Infrared Spectra (FTIR) images. The proposed model integrates three deep learning models including ResNet101, EfficientNetB0, and Wavelet Scattering transform (WST) to extract several features from FTIR. Then the obtained features were fused in conjunction with standard statistical feature extraction. It followed by a subsequent classification phase that employs a Convolutional Neural Network (CNN) architecture, which demonstrates high accuracy in classifying the infrared spectra images into six different classes of ligands and their metal complexes. During the training phase, the network’s weights are iteratively updated using the Adam optimization algorithm. This model addresses the challenge of small and imbalanced datasets through an image oversampling process. Using random over-sampling technique, it enhances the training process and overall classification performance. The extracted features were analyzed using t-distributed Stochastic Neighbor Embedding (t-SNE) to visualize high-dimensional data in two dimensions. The results of the proposed model show high classification accuracy of 0.91%, low error rate of 0.08%, a sensitivity of 0.89% and a precision of 0.89%, false positive rate of 0.01%, F1 score of 0.89, Matthews Correlation Coefficient of 0.87 and Kappa of 0.68.