| In recent years,with the continuous development of the society,the scale of the world's image data is also explosive growth.How to extract useful information quickly and effectively from large amounts of image data has become a research hotspot in the field of computer vision.Image classification and target detection are one of the most important problems in the field of computer vision.At the same time,they are also the basis of other high level vision research such as image segmentation,video target tracking and human behavior analysis.Based on this,the algorithms of image classification and target detection in complex scenes are studied.The main works of this thesis include the following two parts:(1)The theory of image classification based on the bag of words is studied in detail,and the feature coding of the image classification is systematically expounded.Then,the sparse coding algorithm in image classification is emphatically studed.In order to overcome the shortcomes of previous image classification algorithms based on the sparse coding,such as single feature,no spatial structure information of the image and no topological structure information.The multi-scale feature fusion based Hessian sparse coding algorithm is proposed.First,the image is divided into sub-regions with multi-scale spatial pyramid,the Histogram of Oriented Gradient and Scale-invariant feature transform are effectively merged in each subspace layer.Then,the Hessian energy function is derived in detail and the Hessian energy function is introduced into the traditional sparse coding target function as a regular term.Finally,the support vector machine is used for image classification.Experimental results on several recognized databases show that our algorithm can get higher classification accuracy compared with existing methods.(2)In order to improve the performance of the classification and object detection system,the research of deep learning is discussed.A new convolution neural network is proposed for multi-target detection after systematically studying the Faster-RCNN and Google Net detection framework.The main work is to replace the traditional multiple layer extraction feature parts in Faster-RCNN using the inception module of GoogLeNet.The detailed implementation of the whole detection framework is provided.Finally,the complex vehicle recorder image data is used for experimental verification,the results prove the validity of the model. |
