Hyperspectral imaging for green pepper segmentation using a complex-valued neural network

Abstract

Background:Green pepper segmentation is an intractable issue due to its similar color with large amounts of green leaves as background. For the above reason, many existing works incorporate hyperspectral imaging as a powerful tool with richer information than RGB components. However, these methods are all under real values. Objective:This paper mainly investigates the possibility of leveraging a complex-valued neural network(cNN) for hyperspectral classification problem, to get better results for green pepper automatic picking in agriculture. Method:Firstly, we obtain several hyperspectral images in the greenhouse and make pixel-wise datasets as hyperspectral input for the following classification network. Secondly and most importantly, we make a novel attempt by leveraging a small cNN as classification module for binary output, which may be better suited for wave propagation problems. Results:Experimental results have showcased that cNN can generate illumination-robust and clearer outputs with less noise. As for the quantitative results, the outcomes of cNN outperform those from a real-valued one under the same configuration by 3.48 percents in terms of accuracy.