Research on the intelligent detection model of plant diseases based on MamSwinNet

TL;DR

This paper introduces MamSwinNet, a new model for detecting plant diseases that improves accuracy and efficiency compared to existing methods.

Contribution

The novel MamSwinNet model combines efficient token refinement and selective perception modules for better disease detection.

Findings

MamSwinNet achieves high F1 scores of 79.47%, 99.52%, and 99.38% on three plant disease datasets.

The model reduces parameters by 52.9% compared to Swin-T and has a low computational cost of 2.71GMac.

The model enhances spatial and channel feature modeling for accurate disease detection.

Abstract

Plant diseases pose a severe threat to global agricultural production, significantly challenging crop yield, quality, and food security. Therefore, accurate and efficient disease detection is crucial. Current detection methods have clear limitations: CNN-based methods struggle to model long-range dependencies effectively and have weak generalization abilities. Transformer-based methods, while adept at long-range feature modeling, face issues with large parameter sizes and inefficient calculations due to the quadratic complexity of the self-attention mechanism in relation to image size. To address these challenges, this paper proposes the MamSwinNet model. Its core innovation lies in: using the Efficient Token Refinement module with an overlapping space reduction method, relying on depthwise separable convolutions designed with “stride + 3” convolution kernels to expand the image block overlap area and fully preserve boundary spatial structure. This generates high-quality tokens and converts them into a fixed number of latent tokens, reducing computational complexity while maximizing the retention of key features. It integrates the Spatial Global Selective Perception (SGSP) module and the Channel Coordinate Global Optimal Scanning (CCGOS) module. The SGSP module uses a dual-branch structure (the spatial modeling branch introduces 2D-SSM to scan four directions for capturing long-range dependencies, and the residual compensation branch supplements features to prevent loss; the two branches are combined using Hadamard product to enhance spatial detail modeling). The CCGOS module combines channel and spatial attention by embedding positional information through global average pooling in the height and width dimensions, using the Mamba block for channel-selective scanning and generating an attention map, enabling precise association of key channel features like color with spatial distribution. Experimental results show that the model achieves F1 scores of 79.47%, 99.52%, and 99.38% on the PlantDoc, PlantVillage, and Cotton datasets, respectively. The model has only 12.97M parameters (52.9% less than the Swin-T model) and a computational cost as low as 2.71GMac, significantly improving computational efficiency. This study provides an efficient and reliable intelligent solution for large-scale crop disease detection.

Figures