Research On Image Clustering And Detection Algorithm

One important carrier of exchanging information is image in daily life.Research on the image content is referred to as image analysis.The application of image analysis involves various fields such as computer vision,machine vision,signal processing and digital geometry.Therefore,it is of great theoretical value and practical significance to study image analysis algorithms.In this thesis,we studied two important algorithms in image analysis,namely clustering and object detection.From the viewpoint of engineering,clustering and object detection are key technologies of image analysis,and have been widely used in image retrieval,medical imaging,multimedia communication and entertainment.From the viewpoint of theoretical study,clustering is a typical unsupervised learning problem while object detection belongs to supervised learning task.Therefore,researches of these two algorithms have great social value and academic value.There are following contributions:Firstly,GNMF is based on k-NN graph which is known to be sensitive to outliers.In addition,it suffers from the scale transfer problem.In order to defeat the preceding problems,we propose a novel method which jointly incorporates an ingenious Robust Graph and Reconstruction-based Graph regularization into NMF(RG~2NMF)for image clustering.RG~2NMF exploits the learning fashion to generate the robust graph and meanwhile employs the reconstruction regularization to stabilize the objective of GNMF.Experiments of image clustering illustrate the effectiveness of RG~2NMF compared with the baseline methods in quantities.Secondly,YOLO has high positioning error and poor generalization ability for objects with small scales or unusual aspect ratios which makes the accuracy slightly lower than the most advanced level.To address the problem above,we present a new YOLO algorithm based saliency and multi-scale object detection(SMSYOLO).SMSYOLO firstly introduced the idea of multi-scale object detection,matched the objects of different scales on multiple feature maps,and used the deep residual network as the basic network to extract more powerful features.It also used the saliency priori knowledge to improve the positioning accuracy.Experiments of object detection illustrate the effectiveness of SMSYOLO compared with the baseline methods in quantities.