An enhanced MobileNetV4 architecture with CBAM for accurate classification mushroom toxicity

The morphological resemblance between edible and toxic mushroom species presents significant public health challenges, contributing to thousands of annual poisoning incidents globally through visual classification errors. This study proposes an enhanced MobileNetV4 architecture integrating Convolutional Block Attention Module (CBAM) with optimized reduction ratio for accurate mushroom toxicity classification on mobile devices. Using a dataset of 2,000 mushroom images across two classes (edible and poisonous), we systematically evaluated six model variants combining MobileNetV4 Small/Large with CBAM at different reduction ratios. Results demonstrate that MobileNetV4 Large+CBAM 16 achieves 96.90% accuracy with only 13.27M parameters, while maintaining real-time inference speed of 29.7 ms per image. The integration of CBAM with reduction ratio 16 effectively enhances feature representation for distinguishing morphologically similar species, outperforming baseline MobileNetV4 and existing lightweight architectures. This work enables practical deployment of accurate mushroom toxicity classification systems on resource-constrained devices for food safety applications.