Abstract
Affected by various environmental factors, citrus frequently suffers from diseases during growth, posing significant obstacles to agricultural development. This paper proposes a novel method for identifying and classifying citrus diseases. We first design an image enhancement method, HFLF-MS, based on the MSRCR algorithm and Laplacian-optimized homomorphic filtering to highlight disease characteristics. Secondly, we introduce DS-MENet, a neural network based on the DenseNet-121 backbone. DS-MENet utilizes depthwise separable convolutions to reduce parameters, the ReMish activation function to improve robustness, and a multi-channel fusion (MCF) module to enhance feature extraction. Experimental results using 10-fold cross-validation show an average classification accuracy of 95.02% even on noisy datasets, demonstrating the system's feasibility for real-world agricultural applications.
Introduction
Citrus production is a vital component of global agriculture. However, the industry faces constant threats from diseases such as canker, scab, and anthracnose. Traditional manual inspection is labor-intensive and error-prone. While automated computer vision systems offer a solution, real-world field conditions—characterized by uneven lighting, motion blur, and complex backgrounds—often degrade image quality, making accurate classification difficult.
Furthermore, many citrus diseases exhibit high visual similarity, requiring models that can capture fine-grained features while remaining efficient enough for deployment on mobile devices.
Methodology
The proposed system follows a two-stage pipeline: image enhancement followed by deep learning-based classification.
HFLF-MS Image Enhancement
To address poor image quality, the authors developed the HFLF-MS (Homomorphic Filtering optimized by Laplacian - Multi-Scale Retinex with Color Restoration) method. This approach combines MSRCR for color restoration and brightness balancing with Laplacian-optimized homomorphic filtering to sharpen edges. This ensures that the subsequent neural network receives high-contrast images with clearly defined disease features.
The DS-MENet Architecture
The core classification model, DS-MENet, is built upon the DenseNet-121 architecture with three primary optimizations:
- Depthwise Separable Convolutions: By replacing standard convolutions, the model reduces its parameter count by approximately 85%, facilitating faster inference.
- ReMish Activation Function: A hybrid of ReLU and Mish, ReMish prevents "neuron death" by allowing a small gradient for negative values while maintaining the efficiency of ReLU for positive values.
- Multi-Channel Fusion (MCF): This module is integrated into the Dense Blocks to group feature maps and apply attention mechanisms, helping the model focus on the most discriminative disease characteristics.
Experimental Results
The model was trained and validated on a dataset of 9,258 images across five categories (Healthy, Canker, Scab, Black Spot, and Anthracnose).
- Accuracy: DS-MENet achieved a peak average accuracy of 95.25%.
- Robustness: To simulate real-world sensor noise, researchers added Gaussian and impulse noise to the test set. The model maintained a high accuracy of 95.02%, significantly outperforming standard architectures like ResNet50 and InceptionV4.
- Generalization: The architecture also showed strong performance (94-96%) on non-agricultural datasets such as Stanford Cars and ImageNetDogs, indicating its effectiveness as a general feature extractor.
Conclusion
DS-MENet provides a robust and efficient solution for citrus disease classification in challenging environments. By combining sophisticated pre-processing with an optimized deep learning architecture, the system achieves high accuracy with low computational overhead. Future work will focus on expanding the dataset to include more varieties and optimizing the model for real-time mobile deployment to assist farmers directly in the field.
Summary updated