Research And Implementation Of Image And 3D Shape Retrieval Algorithms Based On Deep Learning

Image retrieval and 3D model retrieval are important research fields of Computer Vision.As with the popularization of image capturing devices(e.g.camera,mobile phone,etc.)and 3D model capturing devices(e.g.depth camera,LiDAR,etc.),the scale of images and 3D models are expanding quickly,which makes retrieval more challenging than ever.As a branch of Machine Learning,Deep Learning is fast developing in recent years and exhibits advantages in extracting features of image,texts,voices,etc.Since features do matter for image retrieval and 3D model retrieval,it is possible to improve the performance of retrieval by utilizing the features extracted with Deep Learning.This thesis aims at importing Deep Learning into the fields of image retrieval and 3D model retrieval by analyzing the special needs of retrieval to innovate the application pattern of Deep Learning.For category-level large scale image retrieval,by adding a special Hashing layer,the network is capable to directly extract Hashing-style binarized features,which preserve a lot semantic information of images and outperform those extracted by traditional Hashing algorithms.For instance-level image retrieval,a combining features method is presented based on the finding that the outputs of different layers have different feature patterns,which could largely increase the useful information in the final features and improve accuracy.Meanwhile,a Convolution Network is adapted with the capability of processing 3D models.By introducing triplet loss into the training stage,the features extracted are more suitable for retrieval.All the methods proposed in this thesis are compared with other related state-of-the-art methods by conducting sufficient experiments,which verifies the feasibility and efficiency of these Deep Learning based methods.