Deep learning paradigms for rice seed classification: a comparative study and novel architectures

Automated classification of rice seeds is essential for quality assessment and sorting in modern agriculture. This paper presents a comprehensive study on rice seed classification utilizing deep learning techniques, particularly Convolutional Neural Networks (CNNs). We reviewed existing research and proposed three CNN architectures: Model1 - A basic CNN architecture; Model2 - Integrating transfer learning with MobileNetV2, and Model3 - An advanced version with additional optimization techniques. A novel dataset including 5000 rice seed images from diverse sources was utilized. Pre-processing techniques, including image cropping and resizing, ensured uniform input dimensions for CNN models. Experimental results illustrate the efficacy of the proposed models. Model3 - Incorporating various optimization techniques, achieved a peak accuracy of 95%, surpassing its predecessors. Comparative analysis revealed Model3’s higher accuracy, recall, and accuracy scores, highlighting its robustness in accurately classifying rice seed varieties. In conclusion, this study emphasizes the significance of deep learning techniques in rice seed classification, offering promising solutions for automating seed quality assessment and enhancing agricultural practices.