SCLFormer: A synergistic convolution-linear attention transformer for hyperspectral image classification of mechanical damage in maize kernels

Abstract

Classifying mechanical damage in maize kernels using hyperspectral imaging is crucial for food security and loss reduction. Existing methods are constrained by high computational complexity and limited precision in detecting subtle damages, such as pericarp damage and kernel cracks. To address these challenges, we introduce a novel algorithm, the Synergistic Convolution and Linear Attention Transformer (SCLFormer). By replacing traditional softmax attention with linear attention, we reduce computational complexity from quadratic to linear, thereby enhancing efficiency. Integrating convolutional operations into the encoder enriches local prior information for global feature modeling, improving classification accuracy. SCLFormer achieves an overall accuracy of 97.08 % in classifying maize kernel damage, with over 85 % accuracy for cracked kernel and pericarp damage. Compared to softmax attention, SCLFormer reduces training and testing times by 355.27 s (16.86 %) and 0.71 s (23.67 %), respectively. Additionally, we propose a modular hyperspectral image-level classification framework that can integrate existing pixel-level feature extraction networks to achieve classification accuracies exceeding 80 %, demonstrating the framework's scalability. SCLFormer, serving as the framework's dedicated feature extraction component, provides a robust solution for maize kernel damage classification and exhibits substantial potential for broader spatial-scale applications. This framework establishes a novel technical paradigm for hyperspectral image-wise classification of other agricultural products.