Multi-Task Learning for Ripeness Prediction of Apple Fruits

Apple is among the most consumed fruits in the world, characterized by a great number of varieties and because of that is available all year round. To aid the retailers in increasing their profit and reducing waste, an accurate and reliable method that sorts apples by shelf-life and ripeness, but doesn't damage the fruit, would be beneficial. Such a system will assist the seller in making a decision about the price of the fruit, based on the shelf life. It will take decisions based on different criteria such as storage conditions, variety, and size. This master study presents a novel method that makes use of two non-destructive testing methods: Visual imaging and Near-Infrared images. The acquired images from both techniques are fed in a deep convolutional neural network for fruit ripeness and shelf-life prediction. The adopted approach is a heterogeneous multitask learning framework, jointly trained on regression and classification tasks. Based on the input data and different preprocessing techniques of thermal images, five different feature extraction modes are presented. The analysis of the results obtained after training different neural network architectures will show that the joint learning of both tasks achieves better performance compared to learning each task individually.