A novel lightweight model integrating convolutional neural network and self-attention mechanism for corn seeds quality image recognition

In the field of agriculture, accurate identification of seeds quality based on artificial intelligence algorithms is revolutionary for enhancing crop yields and qualities. The adoption of convolutional neural networks (CNNs) with scalability and self-attention mechanisms with robustness greatly improves the accuracy of seeds quality recognition. However, CNNs rely mainly on local receptive fields, making it challenging to capture long-range spatial dependencies, while self-attention mechanisms are computationally intensive and consume significant computational resources to achieve effective training. To address these limitations, a lightweight, efficient, and high-accuracy model architecture (denoted as LightMCS) is proposed, which consists of an efficient self-attention mechanism (ESA), partial convolution (PConv), and context broadcasting (CB). The LightMCS gradually extracts and refines features hierarchically. Notably, the PConv and ESA efficiently learn local-global representations, and the CB improves the model's capacity and generalization ability. Additionally, training acceleration techniques are employed to further enhance accuracy and reduce training time. The comparable results show that the proposed LightMCS demonstrates a higher classification accuracy (84.12 %) in the quality recognition task of corn seeds. More importantly, LightMCS has only 3.66 million parameters and a total training time of 243 min. The LightMCS achieves a balance between lightness and efficiency while improving classification accuracy. Thus, the proposed LightMCS provides an effective solution for the accurate identification of corn seeds quality, offering a promising tool in agricultural fields where accurate sorting of defective seeds scenarios is required. • A novel lightweight model for accurate classification of corn seeds is proposed. • Optimized self-attention mechanism reduces computational complexity. • The proposed model efficiently learns local-global representations. • Training acceleration reduces training time and further improves accuracy.