Abstract
Accurate field identification of strawberry diseases is hindered by complex backgrounds and subtle inter-class differences. We propose the Class-Attention-based Lesion Proposal Convolutional Neural Network (CALP-CNN), which uses a cascade architecture to locate main lesion objects and propose discriminative details. Achieving a 91.96% F1-score on a self-built dataset, CALP-CNN outperforms state-of-the-art methods by focusing on fine-grained features while filtering environmental noise.
Introduction
Identifying strawberry diseases in real-world field conditions is significantly more challenging than in controlled laboratory settings. Standard computer vision models often struggle with two primary issues: complex background interference (soil, mulch, weeds) and subtle differences between disease classes that may only be distinguishable by minute textures within a small lesion. Existing methods often require labor-intensive bounding box annotations to train effectively; however, the CALP-CNN achieves high-precision recognition using only image-level labels.
The CALP-CNN Architecture
The core innovation is the Class-Attention-based Lesion Proposal Convolutional Neural Network (CALP-CNN). This model employs a three-scale cascade architecture to refine its focus:
- Scale 1 (Raw Image): Provides a coarse-grained view of the entire scene.
- Scale 2 (Main Lesion Object): The Class Object Location Module (COLM) generates a Class Response Map (CRM) to identify the most relevant pixels for a specific disease, cropping the image to focus on the primary lesion and discarding background noise.
- Scale 3 (Lesion Details): The Lesion Part Proposal Module (LPPM) identifies the five most informative sub-patches within the cropped lesion to capture fine-grained textures.
A critical component is the inter-scale pairwise ranking loss, which mathematically forces the model to have higher confidence in its zoomed-in predictions than in its raw image predictions, ensuring the "zoom" branches learn truly discriminative features.
Dataset: The Strawberry Common Disease Dataset (SCDD)
To train and evaluate the model, researchers introduced the Strawberry Common Disease Dataset (SCDD), containing 3,411 high-quality images across 11 disease classes (such as leaf scorch, gray mold, and anthracnose) and a healthy category. These images were collected from fields in ChangFeng County, China, and supplemented with web-scraped data to ensure diversity in angles, focal lengths, and lighting conditions.
Key Results and Discussion
CALP-CNN achieved an overall accuracy of 92.56% and an F1-score of 91.96%. It significantly outperformed the sub-optimal baseline, MMAL-Net, by 6.52% in F1-score and showed a 9.49% improvement over a standard ResNet-50 backbone. Ablation studies confirmed that the background-filtering "Object" branch (Scale 2) was the most vital component, as removing it led to the largest performance drop.
The model proved particularly effective at distinguishing between diseases that appear nearly identical in early stages, outperforming standard backbones by nearly 6% in these difficult scenarios. This success is attributed to the class-aware attention mechanism, which was found to be 5.57% more accurate than standard class-agnostic saliency methods.
Conclusion
CALP-CNN provides a robust solution for field-based strawberry disease identification. By utilizing class-aware attention to "smart zoom" into discriminative lesion details, the model effectively overcomes the noise of agricultural environments. Future work will focus on improving performance for very small lesions and reducing the computational complexity of the multi-branch architecture.
Summary updated