Automatic Blossom Detection in Apple Trees using Deep Learning

Abstract

Overcropping in fruit trees results in decreased fruit size, poor fruit quality, biennial bearing, and reduction in productive life of orchards. Although flowers and fruits are removed/thinned naturally, they require additional thinning for commercial grade fruit production. Integration of machine vision system in mechanical/chemical thinning facilitates automated selective blossom thinning. The primary requirement for automating blossom thinning is to estimate the blossom density in apple trees under varying background and lighting conditions. In this work, we implement Mask-RCNN algorithm to perform instance segmentation of apple blossoms. Different image augmentation techniques were implemented and their impact on blossom detection were assessed. Experiments were conducted to achieve optimal values of hyperparameters of the deep learning network during the training. Implementation of image augmentation was crucial to reduce validation loss and improve detection accuracy of segmentation algorithm. The proposed system achieved average precision (AP) of 0.86 in detecting blossoms in test dataset previously unseen by the network.