Multitask learning approach for fruit ripeness prediction using a dual band thermal camera

Multi-Task Learning allows to comprehend multiple tasks based on a single neural architecture by making the assumption that the tasks are related. We propose a multi-head architecture that allows to classify melon fruits into ripeness classes by taking images based on dual bandaid-and longwave infrared (IR) thermography. To study the ripening labels based on sugar content, firmness and aroma are studied. We are able to show that multi-task learning delivers superior results for multimodal and multi-label situation when different tasks exhibit a positive transfer