Image Segmentation Based on Deep Learning Features

Abstract

Image segmentation is an important technique in image analysis. Existing methods in image segmentation rely on an artificial neural network to extract the feature of the image. In this study, we propose an image segmentation method based on deep learning features and community detection. We propose the use of a pre-trained convolution neural network (CNN) to extract the deep learning features of the image. The deep CNN is trained on ImageNet dataset and transferred to image segmentations for constructing potentials of super-pixels. We first convert the original image from the pixel level to the region level based on Simple Linear Iterative Clustering super-pixels and then aim at each superpixel region to extract the deep learning features. We combine the color features and deep learning features of the superpixel region. The weights of deep learning features and color features are subsequently adjusted using a random walk algorithm to construct a new similarity matrix. We conduct community detection based on a similarity matrix. To automatically identify the number of image segmentation, we use modularity Q in order to determine the optimal number of associations. To illustrate the effectiveness of our proposed method, we evaluate the BSDS300 dataset and compare the technique with several other wellknown image segmentation methods. The segmentation experiments conducted on different images show that our proposed image segmentation algorithm outperforms other methods.