Apple detection and instance segmentation in natural environments using an improved Mask Scoring R-CNN Model.

Area of Science:

• Computer Vision
• Agricultural Technology
• Machine Learning

Background:

• Accurate apple detection and segmentation are crucial for yield estimation and harvesting.
• Challenges include variable lighting, occlusions, and color changes, hindering automated analysis.

Purpose of the Study:

• To develop an advanced Mask Scoring R-CNN model (MS-ADS) for robust apple detection and instance segmentation in natural orchard environments.
• To address limitations of existing methods in handling occlusions, varying illumination, and fruit color transitions.

Main Methods:

• Utilized a ResNeSt with a feature pyramid network as the backbone for enhanced feature extraction.
• Modified R-CNN and mask heads by incorporating convolutional layers and a Dual Attention Network to improve bounding box and segmentation accuracy.
• Implemented the MS-ADS model for real-time apple instance segmentation.

Main Results:

• Achieved high performance metrics: 97.4% recall, 96.5% precision, and 96.9% F1 score.
• Obtained a bounding box mAP (bbox_mAP) of 0.932 and instance segmentation mAP (mask_mAP) of 0.920.
• Demonstrated robust detection and segmentation under challenging conditions like occlusion and varied lighting, with an average processing time of 0.27s per image.

Conclusions:

• The MS-ADS model provides accurate and robust real-time apple detection and segmentation in orchards.
• This method effectively overcomes common challenges in agricultural computer vision tasks.
• Lays the groundwork for automated yield estimation, harvesting, and growth monitoring systems.