Image Segmentation With Superpixel

Segmentation of an image has been intensively studied over the previous years in computer vision for object detection application and other image processing tasks that need high segmentation accuracy.Although different segmentation techniques are available for segmenting an image,superpixel-based segmentation techniques are superior because they provide better grouping cues,ability to capture multi-scale and diverse visual patterns of natural images,and higher efficiency than pixel-based image segmentation methods.The aim of this study is to improve the performance of the existing superpixel segmentation techniques that use the handcrafted features for the purpose of improving the performance of the existing superpixel-based image segmentation algorithms.A superpixel segmentation method that utilizes extracted deep pixel features from a pre-trained convolutional neural network(CNN)was proposed.The proposed method adds the deep pixel features together with the position and color information of the pixels.It was observed that the performance of the proposed method could be enhanced noticeably by applying better initial seed points.Consequently,a step from k-means was incorporated to measure the location of the initial seed points.After this process,the active search method was applied to ensure each pixel correspond to the appropriate seed.Moreover,based on the proposed superpixel segmentation algorithm,a spectral clustering algorithm was proposed to merge the generated superpixels and results in image segmentation outcome.The method was proposed to overcome the shortcoming of the spectral clustering and increase image segmentation quality.It involved superpixel generation,replacing the pixels in the given input image with superpixel,and then computing the similarity matrix using a weighted undirected graph by representing each superpixel as a node.Finally,the superpixels were clustered to generate the segmentation of an image.Segmentation results on BSDS500 dataset demonstrate that the proposed method achieved promising result between efficiency and segmentation quality in comparison with the existing state-of-the-art image segmentation techniques.The qualitative results also showed that the proposed method produced more perceptually satisfying segmentations.Therefore,the proposed technique can be used efficiently for image segmentation.