COMPARATIVE ANALYSIS OF DEEP LEARNING TECHNIQUES AND ATTENTION MECHANISMS FOR MUSHROOM DISEASE DETECTION

Abstract

Deep learning (DL) has transformed image-based diagnosis by identifying learning discriminative features from complex agricultural images automatically. However, it remains challenging to select an appropriate deep learning architecture for mushroom disease detection due to the subtle differences in the visual characteristics of mushroom disease categories and the limited availability of annotated mushroom datasets. In this paper, the authors consider the most popular CNN-based architectures including CNN, AlexNet, VGG16, VGG19, ResNet50, DenseNet121, InceptionV3, and Vision Transformer (ViT), alongside recent attention mechanisms such as Channel Attention (CA), Spatial Attention (SA), SE-Net, Convolutional Block Attention Module (CBAM), Self-Attention, and Transformer Attention and provide a detailed comparative analysis of these architectures. All methods are systematically evaluated in terms of their architectural characteristics, feature learning capability, computational complexity, pros and cons, as well as their suitability for mushroom disease classification. The comparative evaluation demonstrates that DenseNet121 is the most appropriate backbone architecture because of its dense feature connectivity, efficient gradient propagation and effective parameter utilization. The analysis presented will furnish a foundation for developing robust, efficient, and interpretable DL methods for automated mushroom disease detection.