The Image Classification Algorithm Based On The Fractal Theory

With the development of multimedia technology,image is more and more important in the Internet.It becomes a topic of concern to the rapid and effective organization of such rich and intuitive information.Fractal theory can be used to describe and analyze the image texture and spatial information,so it could be considered as the characteristic of the image.In this paper,we propose a series of algorithms based on fractal theory to extract texture characteristic and the experiments are carried out to evaluate the performance of the algorithms.1.Fractal dimension(FD)reflects the intrinsic self-similarity of an image and can be used in image classification,image segmentation and texture analysis.Based on the different fractal theory,the fractal dimension of objects can be computed in many ways.Because of the simple calculation and clear theoretical foundation,the differential box-counting(DBC)method is a common approach to calculating the FD values.However,the traditional method has a large fitting error and ignores the change of the edge.In this study we proposed an improved DBC-based approach to optimizing the performance of the method in the following ways: reducing fitting errors by decreasing step lengths,considering under-counting boxes on the border of two neighboring box-blocks and making better use of all the pixels in the blocks while not neglecting the middle parts.The experimental results show that the fitting error of the new method can be decreased to 0.012879.The average distance of the FD values is decreased by 16.0% in the divided images and the average variance of the FD values is decreased by 30% in the scaled images,compared with other modified methods.The results show that the new method has a better performance in the recognition of the same type of images and the scaled images.2.In recent years,the local descriptor attracted wide interests of researchers,among which LBP-like methods are the representative ones.The local binary pattern(LBP)method won the wide attention of researchers because of the significant advantage of rotational invariance and scale invariance,but all the LBP-like methods have two significant drawbacks.Firstly,all of these methods are sensitive to the noise.The texture features extracted from the image would have a dramatic change even when the image is added with only a few noises.What's more,the methods scalarly quantify neighborhood pixels into binary formats,which results in a lot of information losing.We proposed a texture analysis and classification method combining self-similarity and LBP complex networks,but it is not strictly a LBP method.It could express the vectorized relationships between the center pixel and itsneighboring pixels,which overcomes the disadvantage that LBP is sensitive to noise.The experimental results show that the proposed method has a good classification effect in the segmented images,the rotational images and the noisy images.Most of all,it has a good discrimination power in the practical image,which is the best result of the listed methods with relatively short feature length.3.Human Epidermoid Cancer Cells(Hep-2)provides a effective technique for the analysis of antinuclear antibodies.But there is non-reproducible as a result of the subjective technology.An image classification method is proposed to overcome the drawback in the paper.The 27-size Improved Complete Local Binary Pattern Magnitude(ICLBP_M)descriptor,22-size Morphological Feature Descriptor and 13-size Pixel Difference Feature Descriptor,totally 62-size descriptors,are obtained and used as texture features.The ICLBP_M descriptor could describe the self-similarity and the grey-value difference of the image.The morphological feature descriptor is used to represent the connected components in the image,and the pixel difference descriptor is used to describe the statistical characteristics of the pixel difference at different scales.The Multiple Kernel Learning(MKL)is firstly applied to the recognition of Hep-2 cell image.The experimental results show that MKL gives full play to the potential and makes up the weakness of each descriptor.The overall accuracy of the proposed method achieves 66.49% with the feature length of only 62,which is comparable with other methods.