PDC: Pearl Detection with a Counter Based on Deep Learning.

Area of Science:

• Computer Vision
• Machine Learning
• Agricultural Technology

Background:

• Manual counting of valuable sea pearls is time-consuming and risks damage.
• Non-contact, high-precision pearl detection and counting (PDC) is crucial for commercial production.
• Existing methods lack the accuracy and efficiency required for delicate pearl handling.

Purpose of the Study:

• To evaluate nine object-detection models for non-contact pearl detection.
• To identify the most effective model for accurate and rapid pearl counting.
• To demonstrate the superiority of a proposed Faster R-CNN ResNet152 model for pearl recognition.

Main Methods:

• Comprehensive evaluation of nine object-detection models.
• Implementation and testing of Faster R-CNN with ResNet152, pre-trained on a pearl dataset.
• Performance metrics including mAP@0.5IoU, mAP@0.75IoU, and inference time were analyzed.
• Comparison with eight other sophisticated object detectors.

Main Results:

• Faster R-CNN with ResNet152 achieved mAP@0.5IoU of 100% and mAP@0.75IoU of 98.83%.
• The model demonstrated a fast inference time of 15.8 ms for counting.
• Total loss decreased to 0.00044, with classification and localization losses below 0.00019 and 0.00031, respectively.

Conclusions:

• Faster R-CNN ResNet152 offers a highly accurate and efficient solution for non-contact pearl detection and counting.
• The model's robust performance makes it suitable for various lighting conditions (natural and artificial).
• This method significantly improves upon traditional manual counting, preserving pearl integrity and enhancing commercial viability.