Research On Clothing Image Retrieval Based On Multi-feature Fusion Of Topology

With the rapid development of e-commerce,the online transaction volume of clothing products is rapidly increasing.So quickly and accurately matching the desired clothing style from massive data has become a key issue in clothing image retrieval.Traditional text-based clothing image retrieval methods have problems such as high cost,time consumption,and strong subjectivity due to the need to annotate images.However,traditional content-based clothing retrieval methods also face issues such as background interference and occlusion.Although deep learning methods can solve these related problems,they also have other shortcomings,such as the need for large-scale data support,and the large number of parameters in complex deep learning network models can lead to a rapid increase in algorithm complexity and time consumption.The application of continuous homology to analyze big data is currently a research hotspot,and this method can make up for the shortcomings of the above methods.This thesis mainly studies the application of continuous coherence method to process image data,extracting topological invariant features of images through several different methods,using multiple distance algorithms for measurement,and then fusing them with color texture features and depth features to achieve retrieval.The main work of this thesis is as follows:(1)The method of persistent homology in topology is used to extract image features.In this thesis,we use the principle of continuous homology in algebraic topology to construct different simple complex structures and calculate image topological characteristics.According to the characteristics of digital images,a method of topological feature histogram is proposed.This method uses a sliding window method to sequentially scan the image.Each window constructs a Rips complex and calculates the topological features of different windows.Finally,a topological feature histogram is formed as the topological features of the image.(2)By using different distance measurement methods to calculate the distance of the extracted topological features,this paper adopts the traditional bottleneck distance and Wasserstein distance algorithms,as well as mapping the persistence graph using Gaussian kernel functions to vectors of the same latitude for measurement.Experiments have shown that under the premise of extracting topological features using the same method,the Wasserstein distance algorithm has high accuracy but takes a long time.The method of mapping vectors using Gaussian kernel functions takes less time,but has lower accuracy.(3)Integrating topological features with other features for image retrieval,this paper fuses the extracted topological features with color texture features and depth features respectively.The selection of color texture features is based on a combination of color histograms and Surf features.The MoblieNetv2 model is used for deep features,which first trains the data and then intercepts the output of the previous layer of the network classification layer as deep features.The experiment shows that the retrieval accuracy has been improved to a certain extent by integrating topological features.