Research On Sketch Retrieval Method Based On Feature Fusion

With the popularity and developments of the Internet in recent years,the number of digital images has increased rapidly,and the method of image retrieval is also changed from text-based to content-based.With the advent of a large number of touch screen devices,a new sketch-based image retrieval method has attracted the attention of researchers.Users can draw freehand sketch on the device easily and use it as the query image.Then the system returns some matched images from a large number of image database.Because of the particularity of the sketch-based image retrieval method,there are still many difficulties.For example,it is necessary to find the appropriate feature extraction method and feature preservation method when we extract features from edge image,and we need to adopt appropriate retrieval method to ensure the accuracy of the results.Based on the above problems,this thesis proposes a feature fusion method for sketch-based image retrieval.In feature extraction method,we propose a Multiple Binary Histogram of Oriented Gradient(MBHOG)feature to obtain the texture features of the image.In retrieval method,we propose a binary image mask to improve retrieval accuracy.We propose a method for reordering results based on color features.Finally,we design and implement a sketch-based image retrieval system,the main content of this thesis are as follows:1)A new feature extraction method:MBHOG.Firstly,we improve a feature based on a Binary HOG(BHOG).We extract two main directions as features in each cell.Secondly,we propose a binary encode method,which converts the gradient values of traditional HOG features into discrete values and replaces them with corresponding binary codes,we can get Discrete HOG(DBHOG)feature.Finally,the feature fusion method is used to obtain the MBHOG features.The two feature extraction and preservation methods can effectively reduce the feature storage space,and retain important features,while binary feature coding method can improve the efficiency of retrieval.2)A new mask as the constraint condition for retrieval.The design of retrieval method always matches the feature extraction method.In this thesis,Hamming distance is used as the calculation criterion when calculating the distance of image.In order to improve the accuracy of retrieval results and the robustness of retrieval methods,we propose a binary mask as the constraint condition of distance calculation,which can also reduce the amount of calculation.Lots of experiments have shown that better results can be obtained by using an image mask.Finally we propose a method to convert a single RGB to a color histogram.When we get retrieval results according to the above method,the results can be reordered according to the color parameter.We cannot get a useful color histogram from single RGB value.In this thesis,an effective method is proposed to obtain the effective color histogram of 4096 dimensions.3)A new sketch-based image retrieval system with feature fusion.We can draw sketches on the interface of the system as input,and the system can return and display the relevant results quickly.