Class-attention-based lesion proposal convolutional neural network for strawberry diseases identification.

Area of Science:

• Agricultural science
• Computer vision
• Plant pathology

Background:

• Strawberry diseases significantly impact crop quality and yield.
• Accurate and timely field disease identification is essential but challenging due to complex backgrounds and subtle inter-class differences.
• Segmenting strawberry lesions and learning fine-grained features can address these challenges.

Purpose of the Study:

• To develop a novel method for accurate field strawberry disease identification.
• To address challenges posed by complex backgrounds and subtle disease variations.
• To improve the fine-grained recognition of strawberry diseases.

Main Methods:

• A Class-Attention-based Lesion Proposal Convolutional Neural Network (CALP-CNN) was developed.
• The CALP-CNN utilizes a class object location module (COLM) to locate lesions and a lesion part proposal module (LPPM) to identify discriminative details.
• A cascade architecture was employed to handle background interference and misclassification.

Main Results:

• The CALP-CNN achieved high classification performance on a self-built dataset.
• Accuracy, precision, recall, and F1-score were 92.56%, 92.55%, 91.80%, and 91.96%, respectively.
• The CALP-CNN outperformed six state-of-the-art attention-based methods, showing a 6.52% higher F1-score than the MMAL-Net baseline.

Conclusions:

• The proposed CALP-CNN is effective for identifying strawberry diseases in complex field environments.
• The method successfully addresses background interference and subtle inter-class differences.
• CALP-CNN offers a promising solution for automated disease diagnosis in agriculture.