| The progress of science and technology promotes the development of digital information age.With the iteration and popularization of digital imaging and storage devices and the Internet,people can quickly obtain,store,and disseminate digital images.Therefore,how to retrieve target images quickly and accurately from a variety of growing digital image libraries has become an enduring research hotspot in the field of computer vision and pattern recognition.Based on this research content,this thesis takes color texture image as the research object,and carries out the related image retrieval research from the two directions of feature fusion and deep learning.Among them,the purpose of feature fusion is to combine different types of lowlevel features to get a more comprehensive middle-level feature for image expression,which makes the feature more discriminant.It is a common method to improve the performance of retrieval system in manual image retrieval methods.However,the features extracted by deep learning method can better represent the deep intrinsic features of images,which is a hot spot in image retrieval research.Based on the above background,this thesis does the following work in color texture image retrieval research based on feature fusion and deep learning:(1)To solve the problem that a single feature cannot well and completely express the information of color texture images,an image feature extraction scheme combining complementary color and texture features is proposed,and an optimal closed similarity measure is used to complete the retrieval of color texture images.In this retrieval method,three channels of data are processed in a non-uniformly quantized way in HSV color space,which is close to human visual characteristics,and color features in a combined histogram are extracted.Then,the Gamma distribution parameter features of the subband amplitude coefficients and the von Mises distribution parameter features of the phase coefficients of the V-channel data are extracted from the multiscale and multidirectional Gabor complex transformation domain as the global texture features,At the same time,Local Neighbor Difference Pattern(LNDP)descriptor,which is more sufficient to represent local texture information,is used to extract the local texture features of V channel data,and the obtained color features,global texture features and local texture features are organically fused.Finally,the optimal similarity measure is adopted,that is,the improved Manhattan distance is used for color features and local texture features,and the corresponding Kullback-Leibler(K-L)distance is used for global texture features composed of two distribution model parameters,respectively,to complete the retrieval of color texture images.The experimental results show that the image feature extraction scheme combining color and texture features can effectively improve the performance of color texture image retrieval system.(2)To solve the problem of insufficient extraction and utilization of image phase features,a new color texture image retrieval method based on global and local(relative)phase features fusion is proposed.In this method,the Pattern of Local Gravitational Force Angle(PLGFA)is used as the local phase descriptor for color texture image retrieval for the first time.Specifically,this method firstly extracts color features in the form of combined histogram in HSV color space close to human visual system by using improved non-uniform quantization;Then,in Gabor complex transform domain,relative phase modeling is carried out for V channel data to extract global texture phase features,and at the same time,PLGFA descriptor is used to extract local texture phase features from gray scale image of RGB images.Finally,these three features are fused,and the corresponding closed K-L distance and the improved Manhattan distance are adopted to obtain the total similarity measure,so as to realize the retrieval of color texture image.The experimental results on four standard datasets combined with two indicators show that the method has better comprehensive retrieval performance than the existing methods.(3)Aiming at the problems that the existing deep learning-based color texture image retrieval methods lack large-scale training datasets and the retrieval performance needs to be improved,a new method based on the Vision Transformer network model is proposed,which combines the construction and expansion of datasets.In this method,a dataset with 75,000 color texture images is first constructed,and transfer learning is performed on the tiny network model of Swin Transformer;Then,the network model parameters are fine-tuned using four classical color texture image datasets as the target dataset,and the model is used for feature extraction;Finally,the related retrieval experiments are completed using the Hamming distance as the similarity measure.The experimental results on four target datasets show that the constructed large-scale color texture image dataset has high practical value.At the same time,the relevant experimental results also show that a reasonable enhancement of training data can help to improve the performance of the network model. |
