Files
Abstract
Accurate detection and segmentation techniques are essential to monitor fruit crop growth and perform yield estimation automatically. Recent research on deep neural networks such as Faster R-CNN and Mask R-CNN has shown promising results on automatic fruit detection and segmentation. However, these networks suffer problems of false positives and missing fruits, while these hard examples often exist in fruit images collected in greenhouses due to leaf occlusion and poor lighting conditions. This thesis proposes to augment Faster R-CNN and Mask R-CNN with an online hard example mining (OHEM) algorithm, resulting in considerably improved detection and segmentation results on our dataset. The Mask R-CNN with OHEM jointly detects and segments tomatoes and demonstrates the best performance with an F1-score of 0.957 for detection and a dice sore of 0.816 for segmentation. Applications of the proposed networks include tomato counting and growth monitoring, suggesting the promise of their future deployment in greenhouses.