Semantic segmentation of diamond image using dilated convolutional neural network

For precise segmentation of diamond images, a semantic segmentation model was constructed based on dilated convolutional neural network. A small-scale data set of diamond images was established. The hyper parameters of the model, including batch size, number of filters and dilation coefficient, were optimized. The segmentation results obtained with dilated convolutional network were compared with those acquired by thresholding method and adaptive thresholding method. The results show that batch size, number of filters and dilation coefficient have important effects on the segmentation ability of the constructed model. It is also found that dilated convolutional network can achieve a recall value of 0.966 at the level of 0.965 precision, which is much higher than those of traditional methods. Especially, it is able to classify effectively the bright spots in diamond images.