PRINCE: Advanced classification algorithm for rice grain recognition in clustered images

Abstract

With the rapid development of agriculture, the number of paddy (Oryza sativa L.) is increasing. However, accurately recognizing the variety of rice grain (dehusked paddy) is a significant challenge due to the occlusion and similarity problems in the image recognition field. To address the rice grain recognition problem in clustered images, we propose a novel precision rice grain identification and classification engine (PRINCE) architecture for high-similarity clustered rice grain images. Specifically, we pioneer the exploration and implementation of the SAM model in rice grain analysis, achieving zero-shot semantic segmentation of clustered rice grain images with diverse morphological masks. Secondly, we design a dual-layer filter (D-Filter), where Filter-I is a threshold-controlled discrete rice grain morphology quantitative analysis method for calibrating the morphological integrity of rice grain masks, and Filter-II is a neural network classifier of rice grain mask images that selects complete rice grain mask images from complex mask data. Finally, we integrate dual migration learning and pre-trained model fine-tuning (D-FTL) to train a classification model that accurately recognizes twelve visually indistinguishable discrete rice grain varieties, achieving a weighted F1-score of 82.29%, Top1 accuracy of 82.238%, and area under the curve (AUC) of 0.99. Extensive experimental results show that the proposed PRINCE architecture outperforms seven existing mainstream classification models in terms of accuracy, precision, and recall. Our research demonstrates practical significance in rice variety identification, cooking parameter optimization, and adulteration detection, establishing a novel framework for intelligent grain assessment and optimal cooking outcomes.