The Potential Application of Innovative Methods in Neural Networks for Surface Crack Recognition of Unshelled Hazelnut

Abstract

In some countries, most hazelnuts are cracked using semi-industrial or hand-crafted machines and marketed as open-shell. In the process of hazelnut cracking, because of the different sizes and shapes of hazelnuts, many hazelnuts leave the cracking machine in the form of a cracked or closed-shell. The presence of cracked or closed-shell hazelnuts reduces the marketability of the product. Therefore, after the cracking operation, the separation of cracked or closed-shells from whole hazelnuts has largely been conducted by visual inspection, which is time-consuming, labor-intensive, and lacks accuracy. So, the purpose of this study was to use the deep convolutional neural network (DCNN) algorithm to classify hazelnuts into two classes: open-shell and closed-shell or cracked hazelnuts. To compare the proposed method with pretrained DCNN models, three models including ResNet-50, Inception-V3, and VGG-19 were investigated. The results of the proposed model (accuracy of 98% and $F1$ -score of 96.8) showed that the proposed DCNN has good capability in predicting hazelnut classes. Compared with pretrained models, because of the small size and simple architecture of the proposed model, this model can be a good substitute for a complex and large model such as Inception-V3. Overall, the results indicate that crack on the hazelnut surface can be successfully detected automatically, and the proposed DCNN has a high potential to facilitate the development of a hazelnut sorter based on surface crack.