CNN-Transformer fusion for improving maize kernel storage Quality: A new approach to reducing storage risks

Abstract

The efficient identification of imperfect maize kernels is of crucial importance for optimizing storage management and ensuring quality preservation. Image-based technologies, known for their rapid and non-destructive characteristics, play a pivotal role in this endeavor. However, these technologies encounter significant challenges due to high inter-class similarity and substantial intra-class variability among maize kernels. To address these challenges, we propose CTNet, an advanced model that synergistically integrates the Convolutional Neural Network (CNN) with the Transformer architecture. This integration is further enhanced by incorporating a Feature Attention Module (FAM) and a DW-Swin Transformer, which collectively facilitate the fusion of local and global features. A Fine-Grained Perception Module (FGPM) is employed to augment the model's sensitivity to subtle imperfections. The model, supplemented with a linear classifier, ensures precise discrimination of kernel quality. CTNet was trained on the comprehensive GrainSpace dataset and achieved 1.35 % higher accuracy than the baseline model, SwinTransformer, on the test set. Compared with other methods, our model demonstrates higher accuracy in identifying defective maize grains and achieves more efficient parameter utilization in its structure. Despite moderate parameters and FLOPs, it enhances recognition performance while maintaining low complexity. As imperfect maize grains directly impact quality and storage safety, the model's effective recognition capability is critical for reducing storage losses and ensuring safety, thus providing strong support for smart agriculture.