Application of deep learning and explainable artificial intelligence (XAI) for detecting red chilli powder adulteration

Abstract

To tackle the challenge of Red Chilli Powder adulteration (RCP), an artificial intelligence (AI)-based framework was proposed using empirical analysis of eight pre-trained two-dimensional convolutional neural network (2D-CNN) models for RCP adulteration detection. Moreover, to enhance the convergence and performance of the proposed architecture, an optimiser AdamClr is integrated with minimum and maximin learning rate of 0.00005 and 0.01, respectively. The RCP is categorised into two classes; C1_PRcP includes pure RCP of Jodhpuri (JP) variety, and Class 2 (C2_ARcP) consists of various natural adulterants (i.e., wheat bran (WB), rice hull (RB), wood saw (WS), three low-grade varieties of RCP at the lowest concentration of 5 %). Additionally, the model which outperforms corresponding architectures is further evaluated using explainable artificial intelligence (XAI) technology. DenseNet_169, trained at BS 64, delivers 97.99 % accuracy for detecting natural adulterants (C2_ARcP) in high-grade RCP (C1_PRcP). The XAI model (Grad-CAM and LIME) explained the accurate adulteration prediction of the DensNet_169 2D-CNN model. The heat map obtained from both XAI models illustrated the significant areas that explained the model's decision-making. The proposed model effectively detects RCP adulteration and its applicability can be enhanced by increasing dataset diversity. Overall, the integrated 2D-CNN-XAI approach holds significant potential to revolutionise quality control and assurance in the food industry.