Deep learning models to detect wax bloom on blueberry fruits from images

Identifying blueberry (Vaccinium corymbosum L.) phenotypes is an important task that can help develop novel cultivars better suited for changing climates and marketing requirements. The presence of traits such as a wax bloom that covers the blueberry fruit is essential since it protects the fruit from decay and fungal infection and extends shelf-life, which is especially important for the export market. Phenotyping complex traits such as bloom is usually done manually and, therefore, is costly. We present a shallow deep-learning model for automatically detecting wax bloom in blueberries by training the model using distillation knowledge, where its loss was computed using the L2 distance between two density functions, representing the student and the teacher. Each density function was modeled using Gaussian Mixtures. We made the comparisons using the following machine learning methods: Support vector Machine (SVM), Random Forest (RF), AdaBoost, and Multilayer Perceptron (MLP). Also, we evaluated Convolutional Neural Networks (CNN) architectures using a tradeoff between classification accuracy and model size. With only 690 parameters, the proposed model achieved an accuracy of 98% and represents a promising model, since it is very close to the best accuracy achieved (99.2%) when using larger models like the VGG16 with more than 134 million parameters. A novel data set of blueberries with and without wax bloom was created as an additional contribution and will be available for research use upon request.