Research On Image Complexity Based On Color Feature

Image complexity has been widely used in many different fields such as automatic target recognition, edge detection, image compression, watermarking, image steganography, image classification and retrieval, cognitive psychology, visual aesthetics and so on. However, regardless of its widely application, there isn't any universal definition for the concept of "image complexity" as well as its description as researchers from different fields often have different understanding towards it.The constitution theory[2] is a new body of knowledge which one can use to explore and describe the universal law of objects. According to the constitution theory, the concept of complexity[2] is used to reflect the richness of the inner components of the object in study. Thus, the constitution theory has provided a new way for the research of complexity.In this thesis, two of the most direct features of images related to color have been taken into consideration, namely: the color species feature and the color spatial distribution feature. In accordance with the constitution theory, two general sets[2] describing the color species feature and the color spatial distribution feature of images have been defined respectively and the definitions of the complexity of color species feature and color spatial distribution feature have also been given. Apart from that, a new way to compute the color complexity of images has been proposed by combining the two different color features.Algorithms in both RGB and HSI color spaces have been designed and implemented to demonstrate the feasibility and effectiveness of the method to compute the color complexity of images which is proposed in this thesis. Also, a large number of color images have been tested in the experiment and the results show that:(1) in RGB color space high consistency has been achieved between the human perception and the complexity of the color species feature of images, which proves that the way to compute the complexity of the color species feature proposed in this thesis is effective to some extent. However, the consistency between the human perception and the complexity of the color spatial distribution feature of images in the test is low because of the inefficiency of the RGB color model in reflecting the human visual perception. In fact, the algorithm implemented in the RGB color space would label much more different kinds of colors than that can be distinguished by human beings. As a result, the accuracy of the color complexity of these images in the test has been affected. In order to find out whether the color complexity computation method is effective or not after reducing the number of different colors distinguished by the algorithm, a color quantization operation called CoQuWeiP[43] has been performed on images first before the calculation of their color complexity. It shows that although both of the complexity of the color spatial distribution feature and the whole color complexity of the images are more consistent with the human perception, the effectiveness of the complexity of the color species feature has been affected because of the information loss during the process of color quantization;(2) However, after adopting the pixel similarity rule proposed by N. Ikonomaki[23] et al. in the HSI color space, the connected regions calculated based on the rule are much more consistent with the human perception, and the number of different colors labeled by the algorithm is significantly reduced. The experiment result demonstrated that within the 40 groups of the experiment images, the complexity order of these images based on the calculation of the algorithm in each group highly consist with the human visual perception and the consistency rate reached as high as 80% in the experiment groups which can prove that the image color complexity computation method proposed in this thesis is feasible and effective.