Developing a CNN-Based Machine Learning Model for Cardamom Identification: A Transfer Learning Approach

Abstract

Cardamom, a globally significant spice renowned for its medicinal properties and culinary applications, faces persistent challenges due to adulteration with lower-quality substitutes. In this research paper, we present a novel approach to address this issue through the development of a CNN-based machine learning model for the identification of cardamom. Inspired by prior successful research in plant species identification, we explore two approaches: building end-to-end CNN model from scratch and leveraging transfer learning by utilizing pre-trained models.

To establish a robust dataset, we curate a collection of different images of cardamoms. Following a 7:3 split for training and testing, we initially evaluate the end-to-end CNN model, which achieves a subpar average accuracy of 52.25% across both classes. Subsequently, we adopt a transfer learning-based approach, three pre-trained models (VGG19, Inception V3, and MobileNet) have been selected and fine-tuned to enhance performance. Comparative analysis reveals that the Inception V3 model exhibits the highest testing accuracy of 97.5%, surpassing the MobileNet model's accuracy of 95%. Considering the requirement for immediate identification and efficient utilization on low computational devices, we select the MobileNet model, which demonstrates favorable time efficiency and optimal resource utilization.

In conclusion, our developed MobileNet model proves capable of accurately identifying cardamom. However, further improvements are necessary, particularly for handling complex or unclear background images. This research contributes to the growing area of spice identification and underscores the importance of practical applications and transfer learning approaches for optimizing the performance of machine learning models.