The Research Of Classification On Digital Halftone Images

Halftoning is a technique of converting the continuous grayscaleimage into0and1gray value, inverse halftoning is the reverse process ofhalftone technology and it can be applied to image sharpening, resizing,color correction, compression and other image processing. Existinginverse halftoning technologies are divided into general inversehalftoning and special inverse halftoning. Special inverse halftoningtechnique requires the prior knowledge of halftone image and can get thebest quality of reconstructed images. So it is very important to obtain thetype of halftone images before reconstructing images.In this paper, the classification problem of halftone images is studieddeeply and the main work are as follows:1) The method of extracting effective sub-block of halftone imagebased on unit information entropy is proposed. This paper present themethod to calculate the unit information entropy, which can eliminateredundant information, and then the unit information entropy was usedfor extracting effective sub-block of halftone image. A halftone image isdivided into several sub-blocks according to a rule and the entropyvalues of M Mneighborhood of each pixel in sub-blocks is calculated.Then a entropy matrix was formed. Lastly, the rule of extracting effectivesub-block is created.2) A improved GLCM is proposed to satisfy feature extraction ofbinary image. As the binary image, traditional GLCM of halftone imagesis ineffective, so this paper set eight directions for each pixel of a halftoneimage, and then take D steps in each direction." xor" operation is used tocalculate the values of related pixels and its sum value is regarded as itsfeature vector of halftone images which act as the input of BP neuralnetwork. Experimental results showed that classification accuracy ofimproved GLCM were better than that of the traditional GLCM.3) Segma second-order statistical moments avoid contradiction thattraditional covariance matrix have high time complexity when featurematching on the riemann space, but only adapt the covariance matrix offull rank. In order to avoid this problem, Segma_g set had be proposed.Eigenvalues and corresponding eigenvectors which were rebuilded to Segma_g set. It was proofed that Segma_g set can also expresscovariance matrix characteristics completely when covariance matrixdeterminant is0. And Segma_g was used for feature modeling of halftoneimages. Experiments showed that Segma_g statistical moments had agood time complexity and classification accuracy.