Faster RCNN based leaf segmentation using stereo images

In the plant cultivation, pruning is a crucial stage to improve plant metabolism. A robot may be employed for efficient pruning. Before leaf pruning, the robot must distinguish leaves from branches. Aligned images of a golden melon plant captured using a Red-Green-Blue-Depth (RGB-Depth) camera were used for modeling. The images were captured using different light intensities, distances, heights, and shooting angles. This study employs the faster region convolutional neural network (Faster R-CNN) method using RGB-Depth Camera. Faster R-CNN was developed in 10 epochs and had two scenarios (Scenarios 1 and 2) using two intersections over union (IoU) threshold values (30% and 70%) for testing. Scenario 1 (IoU threshold 30%) had a better score for aggregate testing data with 70.02% than Scenario 2. The correlation between the intensity and mean average precision score was -0.292. There is no impact on the model for the sample of testing data at every distance and degree. We conclude that using an IoU threshold on region proposal influences the scores.