Blackberry Fruit Classification in Underexposed Images Combining Deep Learning and Image Fusion Methods

Area of Science:

• Computer Vision
• Agricultural Technology

Background:

• Increasing berry production faces challenges with labor shortages and fruit waste.
• Uncontrolled lighting in agricultural settings causes image underexposure, hindering accurate ripeness classification.
• Distinguishing blackberry ripeness is difficult due to their dark color and variable lighting.

Purpose of the Study:

• To automate blackberry ripeness classification under diverse lighting conditions.
• To enhance image quality using fusion methods for improved computer vision analysis.
• To address challenges posed by underexposed images in agricultural computer vision tasks.

Main Methods:

• Developed an algorithm combining visible, enhanced visible, and near-infrared spectral images.
• Employed image fusion techniques to improve input image quality before classification.
• Evaluated performance on underexposed and outdoor images, analyzing fusion metrics.

Main Results:

• Achieved a mean F1 score of 0.909±0.074 without fine-tuning and 0.962±0.028 with fine-tuning on underexposed images.
• Demonstrated up to a 12% increase in classification rates in some cases.
• Confirmed the method's utility in enhancing outdoor images, improving contrast without altering color saturation.

Conclusions:

• Image fusion effectively improves blackberry ripeness classification in low-light conditions.
• The proposed method enhances image quality for computer vision applications in agriculture.
• Weighted fusion offers a viable solution for improving contrast in underexposed vegetation images.