A machine vision based pistachio sorting using transferred mid-level image representation of Convolutional Neural Network

Convolutional neural networks have proved to be prominent in various fields of machine vision and image classification. Although it necessitates a large-scale dataset for promising performance, the mid-level representation of these networks can be exploited for specified tasks with smaller annotated image dataset. To this end, by evaluating the generality-specificity of the desired layer as a feature extractor layer, the parameters of Convolutional Neural Networks learned on massive-size dataset like ImageNet can be transferred to a new model. In this study, the images of different sort of pistachios including trashes have been acquired to feed into a new model using a support vector classifier. The ultimate goal of our machine vision system is to separate the desired open-shell pistachios from other defected pistachios as well as trashes. For image segmentation, we use active contour method to detect objects and form both new images of each object type and their augmented images. Since our dataset is not large-scale compared to ImageNet classes, a feature reduction method is performed after the feature extractor layer of pre-trained Convolutional Neural Network. The results show the better performance of the proposed approach in detection of desired-formed pistachio facing unseen test set of images compared to basic approaches.