Squeeze and Hypercomplex Networks on Leaf Disease Detection

Nazmul Shahadat, Anh Nguyen, Ritika Lama

Abstract: Detecting agricultural leaf disease is critical for crop yield and quality, where deep attention models offer promising solutions over traditional methods. This paper introduces a novel approach utilizing Squeeze-and-Hypercomplex networks (SHNets) to detect and classify leaf diseases. The existing Squeeze-and-Excitation network (SENet) enhances feature representation through channel-wise (all channels) feature re-calibration. Unlike this, Parameterized hypercomplex multiplication (PHM) based hypercomplex dense layer is used to calculate cross-channel correlations across channels. This enhances the network’s representational capacity by adaptively recalibrating cross-channel feature maps and sharing weights among channels. We introduce a novel hypercomplex dense layer to inherit hypercomplex advantages in SE-based attention networks. Moreover, using hypercomplex algebra in network design enables more expressive modeling of inter-channel dependencies, capturing complex patterns in leaf imagery. Our proposed SHNet architecture was trained and evaluated on diverse leaf disease datasets, including disease categories and healthy samples. The experimental results on benchmark datasets unequivocally demonstrate the superiority of our proposed SHNet over the state-of-the-art SENet methods in terms of accuracy and computational complexity. This makes SHNet a highly suitable solution for real-time applications in precision agriculture, where the timely detection and classification of leaf diseases can significantly impact crop yield and quality.

External IDs:dblp:conf/icpr/ShahadatNL24