Research On Indoor Scenes Segmentation Based On DBSCAN Clustering

As a basic problem in the field of image processing and computer vision,image segmentation is the main step in image processing.It has wide application prospects in the fields of medical image processing,image recognition and saliency detection.At present,hardware acquisition equipment is constantly improving and developing.RGB-D image is widely used in scene reconstruction and scene representation because it can provide color and spatial information of scene.Therefore,the research on RGB-D image segmentation has important significance and application value.Although there are many image segmentation algorithms for two-dimensional RGB images,these algorithms all face the same problem:it is difficult to distinguish different objects in the scene which are adjacent and have similar colors.Because these segmentation algorithms for RGB image do not take into account the influence of depth contained in the RGB-D image,they can not get a good segmentation result.As the boundary fitness and pixel similarity of segmentation results can not be guaranteed,it is necessary to study a new segmentation algorithm.To solve the above problems,this paper proposes a RGB-D image segmentation method for indoor scene based on DBSCAN clustering.DBSCAN is a density-based clustering algorithm for spatial data.It can efficiently handle abnormal data and is mainly used for clustering spatial data.Compared RGB images,RGB-D images contain more spatial information and put forward higher requirements for segmentation algorithm.The segmentation effect is more obvious in processing RGB-D images containing complex and irregular shape objects based on DBSCAN algorithm and flooding algorithm.The main work and innovation of this paper are as follows:Firstly,generating superpixels by over-segmentation of RGB-D images and measuring the similarity of two superpixels based on the similarity metric function defined in this paper.Then,the DBSCAN algorithm is used to cluster the superpixels with similar color and geometric information into the same classification.In the clustering process,we restrict the diffusion area to reduce computational complexity.1.Firstly,generating superpixels by over-segmentation of RGB-D images.Based on DBSCAN algorithm,the RGB-D image is segmented based on DBSCAN and flooding algorithms.The new method uses superpixel as processing object and defines a new distance measure based on density-based spatial data clustering algorithm(DBSCAN)to measure the similarity between superpixels.The distance measure function not only takes into account the color information,but also contains the geometric information.2.In order to obtain accurate and regular segmentation boundary,we combine the color and geometry information of RGB-D images and use DBSCAN algorithm to merge the generated superpixels.In the process of superpixel fusion,we use different threshold to control color information and geometric information separately,which takes into account both information and avoids information confusion caused by direct fusion.In this paper,we also propose an automatic threshold determination method.For different RGB-D images,this method can effectively determine the threshold of color distance measure and spatial measure in image classification.3.The algorithm of this paper not only takes superpixel as processing unit,which greatly reduces the number of objects to be processed,but also proposes an effective diffusion strategy,which controls the diffusion region in the 360-degree neighborhood around the seed point.The diffusion region is an irregular smooth enclosed region.Compared with the traditional DBSCAN algorithm whose diffusion strategy is applied to the whole image,our diffusion process can greatly reduce the computational complexity.Finally,a lot of experiments have been done on the RGB-D image database of indoor scenes to prove our algorithm's accuracy and efficiency.The experimental results show that the accuracy and processing speed of the proposed method are better than other methods.The new method can obtain accurate segmentation boundary and edge preservation advantage is more obvious especially in the image segmentation results containing complex and irregular shape objects.