Decision-Tree Based Pixel Classification for Real-time Citrus Segmentation on FPGA

According to Food and Agriculture Organization, Mexico is one of the top five citrus producers in the world. In order to achieve the required quality control to export their products, citrus producers require sorting machines able to classify millions of fruits according to certain characteristics, as their size and color. Computer vision provides image processing tools, as image segmentation, that could be used as first stage in a classification process. Fruit classification must be fast in order to be able to process as much fruits per second as possible. In this paper, an FPGA architecture for image segmentation of orange images based on decision-tree models is proposed. A decision-tree model is proposed as an alternative to global thresholding and adaptive thresholding algorithms. It was observed that in this scenario, global thresholding fails due to the noise produced by the fast moving fruits in a classification line, and adaptive thresholding algorithms are not suitable for real-time applications, because of their high requirements in computing power and memory. A decision-tree model requires less hardware compared to both algorithms. The proposed model can achieve real-time segmentation because it is based on pixel serialization, and not on pixel neighborhood processing. The proposed architecture was implemented in a Spartan-6 FPGA. It runs at 60 fps and attains an accuracy of 97.1% of correct segmented pixels, compared to an offline manual segmentation of the frames.