Examination of Lemon Bruising Using Different CNN-Based Classifiers and Local Spectral-Spatial Hyperspectral Imaging

The presence of bruises on fruits often indicates cell damage, which can lead to a decrease in the ability peel keep oxygen away from fruits, and as result, breaks down walls membranes damaging fruit content. When chemicals are oxidized by enzymes such polyphenol oxidase, chemical reaction produces an undesirable apparent brown color effect, among others. Early detection bruising prevents low-quality entering consumer market. Hereupon, present paper aims at early identification bruised lemon using 3D-convolutional neural networks (3D-CNN) via local spectral-spatial hyperspectral imaging technique, takes into account adjacent image pixel information both frequency (wavelength) spatial domains 3D-tensor input fruits. A total 70 sound lemons were picked up orchards. First, all labeled images (wavelength range 400–1100 nm) captured belonging healthy (unbruised) class (class label 0). Next, was applied each freefall. Then, samples time gap 8 1) 16 h 2) after induced, thus resulting 3-class ternary classification problem. Four well-known 3D-CNN model namely ResNet, ShuffleNet, DenseNet, MobileNet used classify Python. Results revealed that highest accuracy (90.47%) obtained ResNet model, followed DenseNet (85.71%), ShuffleNet (80.95%) (73.80%); over test set. had larger parameter sizes, but it proven be trained faster than other models with fewer number free parameters. easier train they needed less storage, could not achieve error low two counterparts.