Research On Texture Image Retrieval Based On Multi-feature Fusion And Deep Learning

With the development of multimedia technology and the advent of the digital age,the number of digital images in Internet databases has grown exponentially.How to search for the required images from various databases has become an urgent problem to be solved.To solve this problem,this thesis starts with texture image retrieval,and conducts research from the perspective of multi-feature fusion and deep learning.Multi-feature fusion is an important means to improve retrieval performance in the field of texture image retrieval.By fusing multiple complementary features,the ability of feature representation of images can be effectively improved.Due to the powerful learning ability of deep learning for image features,the research of texture image retrieval based on deep learning has gradually attracted the attention of many researchers.In this context,this thesis has done the following work on texture image retrieval based on multi-feature fusion and deep learning:(1)In view of the fact that a single local feature or global feature can not completely represent image information,but the fusion of two or more complementary features can effectively improve the retrieval performance.From the point of view of feature fusion,a texture image retrieval method combining global and local features is proposed in spatial domain and transform domain.In the spatial domain,the local binary pattern value of the image is calculated,and the histogram is established as the feature;in the transform domain,the dual-tree complex wavelet transform is selected to decompose the image,the low-frequency approximate sub-band coefficients obtained after decomposition are selected for Gaussian Mixture Model modeling,magnitude sub-band coefficients are selected for Gamma distribution model modeling,and relative phase sub-band coefficients are selected for von Mises distribution model modeling,The local binary pattern value of the magnitude sub-band coefficient and the improved local Tetra pattern value of the relative phase sub-band coefficient are calculated.According to the influence of different types of features on retrieval performance,the optimized weight coefficient is set for each type of feature,and a new similarity measurement formula is proposed.Experimental results show that the performance of texture image retrieval can be effectively improved by fusing local features and global features.(2)In order to solve the problem of insufficient utilization of directional sub-band information in dual-tree complex wavelet transform domain,a new local texture descriptor is proposed and combined with statistical modeling in transform domain for texture image retrieval.The proposed local descriptor calculates the eight directions of the central pixel by using the relationship between the central pixel and the neighboring pixels in six directions,which is called the local eight direction pattern.In the texture image retrieval system of this thesis,the feature extraction part combines global statistical features and local pattern features.Among them,both the relative magnitude sub-band coefficients and relative phase sub-band coefficients are modeled as wrapped Cauchy distribution in the dual-tree complex wavelet transform domain,and the global statistical features choose the parameters of this model;while the local pattern features respectively choose the local binary pattern histogram features in the spatial domain and the local eight direction pattern histogram features of each direction sub-band in the dual-tree complex wavelet transform domain.On the other hand,the similarity measurement selects matching distances for different features and combines them in the form of convex linear optimization.Experimental results show that the proposed local eight direction pattern can make full use of the directional sub-band information,and the new retrieval method combining local pattern features and statistical features has better performance than the existing retrieval methods.(3)In order to solve the problem of lack of large texture image database in texture image retrieval based on deep learning,a texture image database of more than 80000 texture images is made,and the VGG-13 structure is fine-tuned so that it can be used for texture image training.The texture image database is pre-trained by using fine-tuned VGG-13,and then the neural network structure is Fine Tuning,by using the classical texture image database.Finally,the feature is extracted by using the fine-tuned pre-training structure,and the Euclidean distance is used to measure the distance.Experiments are carried out on three classical texture image databases.The experimental results show that the texture image database has high application value.The fine-tuned neural network structure can be more suitable for texture image retrieval.