Effect of optimisations on CNN performance in identifying adulteration of red chilli powder

Abstract

This study presents a deep learning framework for the detection and classification of adulteration in Red Chilli Powder using image-based analysis. Besides the aim is to analyse the effect of various optimisation techniques to balance the classification performance of the DenseNet models. Hence, two architectures, DenseNet-121 and DenseNet-169, were optimised using three optimisers (namely AdamClr, RMSprop, and AdamW) across three batch sizes (i.e., 16, 32, 64) over 150 epochs. Further, the performance of the models was analysed on two custom datasets: one for binary classification (pure vs. adulterated RcP) and the second for multiclass classification, identifying six natural adulterants at concentration levels 5%, 10% and 15%. The performance was evaluated using standard metrics derived from confusion matrices, while Gradient-weighted Class Activation Mapping visualisations enhanced model interpretability. The DenseNet-169 combined with AdamClr optimizer at batch size 32 achieved the highest binary classification accuracy (98.92 %) and robust convergence. In multiclass classification, the same configuration demonstrated superior generalisation (accuracy: 96.73%). AdamClr consistently outperformed RMSprop and AdamW, offering stability, rapid convergence, and resilience across datasets. RMSprop exhibited batch size sensitivity and instability, while AdamW, delivered occasional competitive results but lacked robustness without meticulous hyperparameter tuning. Overall, the results highlight the critical role of dynamic learning rate strategies and model depth in optimising convolution neural networks performance for food adulteration detection. This framework demonstrates strong potential for real-world applications in spice authentication and food quality assurance.