Research On Multi-scale Feature Fusion And Canny Edge Detection Based Structure Extraction And Texture Filtering Algorithm

Image is one of the main sources for people to perceive the world.As the basic feature of image,structure contains a lot of meaningful information.Therefore,extracting structural information from natural image which includes abundant texture information is a hot topic in the field of image analysis and pattern recognition.Technology of extracting image structure can not only greatly improve the quality of image understanding,but also can be applied to computer vision,such as object recognition,image segmentation,saliency analysis,and so on.It has extensive research and application implications.However,due to the diversity and complexity of the natural texture image,image structure extraction is still a challenge.In addition,texture filtering method in recent years requires removing unnecessary texture details while preserving the integrity of the distinct structure.Therefore,filtering effect largely depends on the quality of the structure extraction of the image.Existing structure detection algorithms are easily interfered by some strong gradient textures in the image,which leads to the difficulty of extracting the saliency structure of the image.In addition,most of the existing texture filtering algorithms tends to process weak gradient texture images,and the texture for strong gradient is invalid.Especially they cannot keep structure and texture smoothly at the same time.In order to solve the existing problems in the above methods,a structure extraction algorithm on the basis of multi-scale feature fusion and Canny edge detection has been proposed,as well as a three-edge texture filtering algorithm.The main research contents and innovations of this paper are as follows:Firstly,we propose a structure identification method based on multi scale feature fusion and machine learning.First of all,this paper analyzes and extracts multi-scale mixing features based on internal variational,interval gradient and Gabor surround suppression,which have higher resolution for strong-gradient textures with different scales.Then,taking into account the correlation between the various features and the redundancy,this paper uses principal component analysis method to reduce the dimensionality of the extracted features.Finally,this paper compares the performance of several different machine learning models by experiments,and eventually finds that the neural network model can get the best classification results.Secondly,we propose a structure detection algorithm that combined with Canny edge detection and structure identification.Based on the results of multi-scale Canny edge detection,to eliminate the texture pixels which are wrongly checked into the structure,this paper uses deburring,outlier removal and texture edge suppression strategy based on structure prediction graph at first.Then,through breakpoint connection strategy,the structure pixels which missed are recovered.Finally,the structural correction strategy is adopted to solve the existing structural offset problem,and finally the finer structure detection results are obtained.Thirdly,on the basis of the result of structure detection,the filter kernel that judges whether to cross the structure is introduced into the bilateral filtering technique framework.A structure-guided trilateral texture filtering algorithm is proposed and different filtering methods are applied to the structure and texture respectively to obtain structure preserving and texture smoothing results.Compared with the existing algorithms,the effect of this algorithm is markedly improved.Finally,we apply the filtering results to detail enhancement,image stylization,image segmentation,seam cutting and other fields,and facts proved that all get good effect.Experimental results show that the proposed structure detection and texture filtering algorithms have achieved some advantages compared with the existing algorithms in the recognition and maintenance of weak-gradient structures as well as the suppression and smoothing of multi-scale strong-gradient textures.The effectiveness of the proposed algorithm is verified.