Zhenyang Chen, Helong Yu, Shaozhong Song, Chunguang Bi, Jiayi Guo, Xi Ling
Japonica rice, a vital staple crop, exhibits considerable quality variations across different varieties, significantly influencing yield, economic viability, and food processing outcomes. Traditional classification techniques often fall short in terms of efficiency and precision. To tackle this challenge, this research introduces a deep learning approach for the effective classification of Japonica rice seed varieties. A dataset was constructed by collecting 200 seeds from each of nine Japonica rice varieties, with a total of 1800 seeds. Image segmentation and data augmentation techniques were applied to build the dataset. During the pre-training phase, fifteen established deep learning models were evaluated, with the ResNeXt50 model emerging as the top performer. Consequently, the ResNeXt50 model was refined through the substitution of the activation function, convolution operation, downsampling method, and the integration of the ECA attention mechanism, resulting in the J-Rice-ResNeXt network model. The proposed J-Rice-ResNeXt model achieved an impressive accuracy rate of 98.81 %, a 7.82 % improvement over the original ResNeXt50 model, while concurrently reducing parameter count and computational demand by 10.279 M and 1.573 G, respectively. Furthermore, the model demonstrated faster convergence and superior fitting capabilities. Experimental findings confirm the efficacy of J-Rice-ResNeXt in accurately classifying Japonica rice seed varieties, positioning it as a valuable tool in rice breeding, quality assessment, and market grading, and providing robust support for the advancement of agricultural intelligence and precision farming. • Construction of datasets by threshold segmentation and data enhancement. • A dynamic learning rate adjustment strategy was proposed. • Improvement of the ResNeXt50 model through four aspects. • Explaining the improvement process with Grad-CAM. • J-Rice-ResNeXt achieves 98.81 % accuracy on the test set.