Abstract
Grape disease is a significant contributory factor to the decline in grape yield, typically affecting the leaves first. Efficient identification of grape leaf diseases remains a critical unmet need. To mitigate background interference in grape leaf feature extraction and improve the ability to extract small disease spots, we developed a novel method for disease recognition and classification. First, Gaussian filters Sobel smooth de-noising Laplace operator (GSSL) was employed to reduce image noise and enhance texture. A novel network designated coordinated attention shuffle mechanism-asymmetric multi-scale fusion module net (CASM-AMFMNet) was subsequently applied. CoAtNet was employed as the network backbone. The CASM-AMFMNet was further utilized to capture and target grape leaf disease areas, reducing background interference. Finally, an Asymmetric multi-scale fusion module (AMFM) was employed to extract multi-scale features from small disease spots. Experimental results showed that CASM-AMFMNet achieved an accuracy of 95.95%, F1 score of 95.78%, and mAP of 90.27%.
Introduction
Grape cultivation is a global industry, but yield is frequently threatened by fungal and bacterial diseases that manifest first on leaves. While computer vision offers a solution for automated monitoring, three major hurdles remain: image quality issues in real-world photos, background interference from vineyard environments (soil, branches, shadows), and scale variance where early-stage disease spots are microscopic and difficult for standard convolutional neural networks (CNNs) to detect after multiple layers of downsampling.
Methodology
The proposed CASM-AMFMNet architecture utilizes a three-stage pipeline to address these challenges, moving from raw image acquisition through specialized preprocessing to final classification.
GSSL Preprocessing
The Gaussian filters Sobel smooth de-noising Laplace operator (GSSL) algorithm grayscales the image, applies high-pass filtering and Sobel operators to create a mask, and uses a Laplace operator to sharpen textures. This ensures the "veined" texture of the leaf is prominent while noise is suppressed.
Network Architecture
- Backbone (CoAtNet): The model uses CoAtNet, which hybridizes local feature extraction of ConvNets with global modeling capabilities of Transformers, alleviating gradient explosion issues.
- CASM (Coordinate Attention Shuffle Mechanism): Unlike standard attention modules, CASM embeds positional data along two spatial directions. It uses group convolutions and a "channel shuffle" to integrate feature information across channels efficiently.
- AMFM (Asymmetric Multi-Scale Fusion Module): To catch tiny spots, this module extracts features at multiple scales using Asymmetric Convolution Blocks (ACB). By replacing standard $3 \times 3$ kernels with $1 \times 3$ and $3 \times 1$ combinations, the model improves robustness to rotations while reducing parameters.
Dataset and Augmentation
The primary evaluation used a self-made grape leaf dataset containing 3,409 original images collected from the Tianlu vineyard in China and online sources.
The dataset includes five categories: Healthy, Black Rot, Black Measles, Leaf Blight, and Downy Mildew. To simulate various field conditions, the dataset was expanded to 15,824 images using vertical/horizontal mirroring, contrast adjustment, and perspective transformation.
Results and Discussion
The CASM-AMFMNet demonstrated superior performance across multiple metrics:
- Accuracy: Achieved 95.95%, a 7.21% improvement over the base CoAtNet model.
- Precision and mAP: Recorded a precision of 96.00% and a mean Average Precision (mAP) of 90.27%.
- Efficiency: The optimized architecture reduced training time significantly, with a training batch taking 270 seconds compared to 586 seconds for the unoptimized model.
The model outperformed ten other state-of-the-art models, including YoloV4 and DICNN, particularly in identifying "Black Rot" and "Leaf Blight," which are often confused due to similar early-stage appearances. Key innovations like Switchable Normalization (SN) in the CASM module allowed the model to dynamically adjust to high variance in outdoor lighting.
Conclusion
CASM-AMFMNet provides a feasible scheme for deep learning to correctly recognize grape diseases during agricultural production. While highly effective, future work will focus on identifying "mixed" infections where a single leaf has multiple diseases and optimizing the model for mobile deployment to provide real-time diagnostic tools for farmers.
Summary updated