Machine Vision for Grape Cluster Quality Assessment

Cluster thinning is a common pre-harvest viticulture practice to reduce crop load towards improved fruit composition and higher wine quality. Crop load reduction is performed hierarchically, removing as a priority the grape clusters that are considered defected, e.g., malnourished, rotten, infected, etc. Automated cluster thinning could reduce human labor and significantly increase the efficiency of the practice. This work presents a machine vision methodology for defected grapes detection. All grape berries in images are detected using a Mask Regional Convolutional Neural Network (Mask R-CNN), reaching intersection over union (IoU) of up to 98.07%. Then, a set of 159 morphological, color and texture features is extracted from the segmented berries based on Wavelet Coefficient method and Local Binary Patterns (LBPs) in CIELAB color space for the L*, a* and b* channels. Sequential feature selection is then performed to derive optimal features. Extracted features are used to train a Random Forest (RF) classifier to classify the grape berries as healthy or defected, using 10-fold cross-validation. The methodology is tested for both red and white grape cultivars reporting a promising quality assessment accuracy.