Research On Image Retrieval Method Based On SURF And Metirc Learning

With the booming progress of digital technology,the popularity of image formation facilities is increasing,the Internet technology is increasingly developed,and the demand for image information is increasing.How to rapidly and precisely obtain useful images from massive network images is an urgent problem faced by computers.Image retrieval technology is to find the similarity relationship between them by extracting image information,use the metric method to fulfil the similarity retrieval of the source image in the image database.As an important part of image retrieval task,measurement method is an essential step in the retrieval process.Finding an appropriate similarity measurement method can effectively improve the retrieval performance.Usually,there will be complex situations such as background blur,occlusion and illumination change in the image imaging process.It is very necessary to study a more robust matching algorithm for similarity measurement.In the research process of image retrieval,it is found that the similarity matching of the extracted features using the measurement method has a good performance in the retrieval accuracy and efficiency.To solve task of retrieval of different images,the following research and work are carried out in this dissertation:(1)This dissertation analyzes the current image retrieval technology,deeply discusses its key links,systematically researches and analyzes the correlation theory of feature selection and similarity metric,including extracting local features and depth features for retrieval,and finally briefly summarizes the important characteristics and basic structure of convolutional neural network and common retrieval hash algorithms,so as to provide theoretical support for subsequent chapters.(2)Aiming at the shortcomings of SURF algorithm,such as large amount of calculation and high false matching rate,this dissertation proposes an improved SURF image retrieval method.Bresenham circle detection principle is adopted to improve the efficiency of feature detection.Simultaneously,the idea of principal component analysis is drawn into to lower the dimensionality of the extracted feature descriptors.Finally,the matching algorithm is improved by combining pre-inspection and false matching elimination.The improved SURF image retrieval method can not only extract stable feature points,but also accurately match image feature points.It is proved that the improved algorithm in this dissertation can effectively improve the running time and guarantee the accuracy of feature matching.The number and distribution of correct matching point pairs are relatively uniform and ideal,and there are almost no wrong matching points.On the public dataset,the algorithm is robust to the changes of image angle,illumination and distance,and the matching accuracy is also improved.The retrieval effect is better in the image retrieval task,but the retrieval effect is not ideal for some fine-grained images in the dataset.(3)For fine-grained image retrieval,this dissertation also proposes an improved retrieval method.Experiments on different datasets have found that traditional retrieval methods make it difficult to distinguish subtle differences between fine-grained images.On this basis,the Res Net50 network is fine tuned and its features are extracted.The asymmetric convolution is used to improve the residual structure,improve the pooling layer to pyramid pooling,increase the hash layer to save the amount of calculation,and introduce the idea of metric learning to add the triple loss to the loss function for training.In the selection of training set,we use fine-grained image data to facilitate the targeted learning parameters of the network.The performance on Corel and Oxford flowers datasets proves that the training accuracy of WRes Net network in this dissertation is higher than that of other feature extraction networks,can obtain highquality features,and achieves better retrieval results on two fine-grained datasets,and has better retrieval performance than other Hash methods.Figure [39] table [11] reference [73]...