Abstract: Classification of grain quality is essential to the agricultural and food processing industries; without this proper identification of the grain types and impurities, it may lead to potentially reduced standards and increased safety risks to consumers. In this project, such comparative analysis has been performed with respect to the two most popular deep learning models-MobileNetV2 and ResNet50-which are applied to grain images for classification. Both models were trained and tested using a custom dataset developed-by the use of various classes of grains-for accuracy, prediction confidence, and class-wise performance. Furthermore, the system will incorporate a user-friendly web interface developed using Streamlit, enabling uploading of grain images and classification results together with visualizations of model confidence. Results indicate the trade-off between lightweight efficiency in MobileNetV2 and rich deeper representation in ResNet50. This work would show the employability of deep learning models into accessible web applications for applied grain inspection tasks while imparting knowledge about the model selection for embedded or real-time scenarios.

Keywords: Grain Classification, Deep Learning, MobileNetV2, ResNet50, Streamlit, Image Recognition, Web-Based Deployment.

| DOI: 10.17148/IARJSET.2025.12634