Click on “Download PDF” for the PDF version or on the title for the HTML version.

If you are not an ASABE member or if your employer has not arranged for access to the full-text, Click here for options.

Abstract. Deep learning methods recently gained much attention in computer vision research due to their superior performance in image classification and object detection. Blueberry internal bruising detection is a significant challenge for the blueberry industry, especially for early bruising detection using noninvasive methods. The main goal of this study was to detect blueberry internal bruising after mechanical damage from hyperspectral transmittance images using fully convolutional networks (FCN). A total of 600 hand-harvested blueberries were divided into 6 groups with two storage times (30 min, and 3–24 h), and three position treatments (stem bruise, equatorial bruise, and non-bruise). A near-infrared hyperspectral imaging system was used to acquire transmittance images from 970 to 1400 nm with 5 nm bandwidth. Images were acquired from two orientations (calyx-up, stem-up) for each fruit. Random forest and linear discriminant analysis were applied to the spectra to select three key wavelengths to generate 3-channal input images. A total of 1,200 3-channal images were evenly and randomly divided to form training and testing sets. Each image was manually labeled with annotated ground truth. Standard fully convolutional networks (FCN-8s) with VGG-16 NET was utilized to train the 600 images. The average intersection over union accuracy was 77.9% for predicting bruised, non-bruised, and calyx end tissues for 600 testing images. The result indicates that blueberry bruising and calyx end can be segmented from blueberry fruit as early as 30 min after mechanical damage using deep learning method. It will help quantify blueberry bruising more accurately in the future work.