Fruit and Vegetable Segmentation with Decision Trees

Objective of our study is to develop a method to quickly and efficiently identify different types of fruits and vegetables that are travelling on a conveyor. Twenty-four different categories of fruits and vegetables, with 80 images per category, are used for training. Segmentation is first performed on the images with two segmentation masks, which are then downsampled using max-pooling to 25% of the original size. The masked images before downsampling are used with local binary patterns and HOG methodologies for extraction of features to get the textures and shapes. PCA performed on downsampled images along with extracted features reducing the number of principal components but is discarded due to too great a loss in accuracy. Finally, these are fed into the classifier to identify the category of fruit or vegetable. Classification is performed using bagged decision trees. The conclusion indicated the high level precision of our proposal along with faster runtime than Inception.