Automatic Parameter Tuning for Adaptive Thresholding in Fruit Detection.

Area of Science:

• Computer Vision
• Agricultural Technology
• Image Processing

Background:

• Fruit detection is crucial for agriculture but challenging due to variable lighting and fruit characteristics.
• Existing methods often struggle with robustness in natural growing conditions.

Purpose of the Study:

• To develop and evaluate an automatic parameter tuning procedure for a dynamic adaptive thresholding algorithm for fruit detection.
• To enhance the algorithm's robustness and adaptability to diverse fruit types and illumination conditions.

Main Methods:

• Developed an automatic parameter tuning procedure for dynamic adaptive thresholding.
• Utilized four color spaces (RGB, HSI, LAB, NDI) with nine thresholds across three illumination levels.
• Trained and tested the algorithm on natural databases of apples, grapes, and peppers using cross-validation.

Main Results:

• Achieved high F-scores: 93.17% for apples and 99.31% for peppers in variable lighting.
• Demonstrated the importance of the tuning process for algorithm generalization across fruits and environments.
• Identified that different color spaces (e.g., LAB for noise robustness) may be superior for specific fruits.

Conclusions:

• The automatic tuning procedure significantly improves the generalization of the dynamic adaptive thresholding algorithm.
• Evaluating multiple color spaces is essential for optimizing fruit detection performance.
• The algorithm shows robustness to lighting variations, threshold changes, and image noise.