Transfer Learning Based Fruits Image Segmentation for Fruit-Picking Robots

Abstract

It is an important prerequisite for a fruit-picking robot to accurately segment and locate the object in fruit images. However, image segmentation by manually selected features or deep learning-based approaches is a troublesome task. It requires a long time and a large number of annotated images for the model to be trained. In this study, transfer learning is used so that the learned parameters of a pre-trained convolutional neural network can be used as the initial settings in the new task. Three networks, Mobilenet_v2, Resnet_v1_50_beta and Xception_65, are used as backbone networks, which were used in the well-known semantic image segmentation model-DeepLab. The proposed transfer learning-based fruits image segmentation not only alleviates the stringent need of a large image dataset, but also saves much time for training. Experimental results show that the Xception_65 based network has the best performance in terms of the segmentation metric of mean intersection over union. A high-precision instance fruits segmentation guarantees subsequent accurate locations of fruit images for fruit-picking robots, which is of great significance for intelligent agriculture.