Learning an optical filter for green pepper automatic picking in agriculture

Abstract

Green pepper automatic picking has been a long-standing challenge in agriculture due to the similar color between green peppers and green leaves. To tackle this intractable problem, we tried to distinguish between them by using hyperspectral information as prior knowledge. As our core insight, a novel optical filter was designed as a pre-processing tool to find valuable wavelengths where peppers differ a lot from leaves. To this end, firstly, the parameters of the optical filter were learned by end-to-end training with a neural network for pixel-wise hyperspectral input. Secondly, the learned optical filter was applied to hyperspectral data to obtain filtering RGB images, which will be sent to further segmentation framework. Thereby, a two-stage method for green pepper segmentation was proposed, and promising results were achieved owing to the incorporation of the optical filter.