Hae-Kwang Kim

Region-based shape descriptor invariant to rotation, scale and translation

Abstract A region-based shape descriptor invariant to rotation, scale and translation is presented in this paper. For a given binary shape, positions of pixels belonging to the shape are regarded as observed vectors of a 2-D random vector and two eigenvectors are obtained from the covariance matrix of the vector population. The shape is divided into four sub-regions by two principal axes corresponding to the two eigenvectors at the center of mass of the shape. Each sub-region is subdivided into four sub-regions in the same way. The sub-division process is repeated for a predetermined number of times. A quadtree representation with its nodes corresponding to regions of the shape is derived from the above process. Four parameters invariant to translation, rotation and scale are calculated for the corresponding region of each node while two parameters are extracted for the root node. The shape descriptor is represented as a vector of all the parameters and the similarity distance between two shapes is calculated by summing up the absolute differences of each element of descriptor vectors. Experimental results conforming to the MPEG-7 shape descriptor core experiments are presented.

Invariant texture retrieval using modified Zernike moments

Abstract This paper presents an effective texture descriptor invariant to translation, scaling, and rotation for texture-based image retrieval applications. In order to find the minimal matching distance between two descriptors, existing frequency-layout descriptors require a lot of distance calculations with every possible combination of scaling and rotation values because they are not invariant to geometrical transformation. To cope with this problem, a new compact descriptor is proposed that is theoretically invariant to such transformations. The proposed descriptor is obtained by first calculating the power spectrum of an original texture image for translation invariance and then the power spectrum image is normalized for scale invariance. Finally, modified Zernike moments are calculated for rotation invariance. The proposed algorithm is simpler and lower than conventional algorithms in terms of the computational complexity. The effectiveness of the proposed descriptor for invariant texture retrieval is shown with various texture datasets by comparing the retrieval accuracy, the descriptor size, and the matching complexity of the proposed descriptor with those of conventional descriptors.

A modified Zernike moment shape descriptor invariant to translation, rotation and scale for similarity-based image retrieval

Zernike moments are used as a shape descriptor for complex shapes such as trademarks that are difficult to be defined with a single contour for similarity-based image retrieval applications. Zernike moments of a given shape are calculated as correlation values of the shape with Zernike basis functions in that all the pixels of the shape regardless of their positions contribute with the same weight to the Zernike moments. The proposed modified Zernike moment descriptor for a shape is obtained taking account of the importance of the outer form of the shape to human perception. The modified Zernike moment descriptor is obtained by first dividing the original shape into two parts of inner and outer regions with a predetermined radius and then calculating the Zernike moment of the outer part and the inner part of the shape. The proposed descriptor consists of Zernike moments of outer and inner parts. Euclidean distance is used for computing the distance measure between two shapes. For perceptual similarity-based retrieval, the Zernike moments of the outer part are used and for exact-matching retrieval, both of the outer and inner Zernike moments are used. Experimentation under various test conditions shows the effectiveness of the proposed modified Zernike moment descriptor.

Translation, scale, and rotation invariant texture descriptor for texture-based image retrieval

A texture descriptor invariant to translation, rotation and scale changes is presented in this paper. A power spectral image is obtained by using the DFT to extract the proposed descriptor for the translation invariance of a given texture. The power spectral image is scale-normalized by a cut-off frequency on the power spectral image that is calculated as the total energy inside a circle with its radius of the cut-off frequency amounts to a predetermined value. Finally, the rotation invariant Zernike moments are calculated on the translation and scale normalized image for a rotation, scale and translation invariant descriptor. The extraction is simple and fast, using fast known DFT and Zernike transformation. The matching is also simple and fast using the Euclidean distance measure between a query and test textures. The effectiveness of the proposed algorithm is shown with various texture databases.

Adaptive Transform Coefficient Scan for H.264 Intra Coding

This paper presents an adaptive transform coefficient scan method that effectively improves intra coding efficiency of H.264. Instead of applying one zig-zag scan to all transform blocks, the proposed method applies a field scan to a horizontally predicted block, a horizontal scan to a vertically predicted block, and a zig-zag scan to blocks predicted in other prediction modes. Experiments based on JM9.6 were performed using only intra coding. Results of the experiments show that the proposed method yields an average PSNR enhancement of 0.16 dB and a maximum PSNR enhancement of 0.31 dB over the current H.264 using zig-zag scan.

Boundary block-merging (BBM) technique for efficient texture coding of arbitrarily shaped object

We present an efficient texture coding method which enhances the coding efficiency of conventional discrete cosine transform (DCT) with padding techniques for arbitrarily shaped objects in object-based video coding where shape information is provided. The BBM (boundary block-merging) technique is applied to the boundary macroblocks of 16/spl times/16 pixels of a VOP (video object plane) which consist of both background and object pixels. A macroblock consists of four subblocks of 8/spl times/8 pixels. For boundary subblocks consisting of object and background pixels, padding is performed in the background region. For a pair of padded boundary subblocks in a macroblock of which alignment belongs to a predefined set, one subblock is rotated 180/spl deg/ and merged into another one if object pixels do not overlap. After merging, the boundary macroblock is coded using the conventional DCT coding. The merging process reduces the number of subblocks to be DCT coded, and high correlation between adjacent subblocks makes the number of DCT coding bits small. Experimentation has been done on various test sequences under different test conditions, and verifies significant coding efficiency improvement: reduction of coding bits for luminance boundary blocks by 5.7-11.9% at the same PSNR values compared with the padding-based DCT without BBM.

Polyphase Downsampling Based Multiple Description Coding Applied to H.264 Video Coding

This paper presents a video coding method that improves error resilient functionality of H.264 with good coding efficiency. The method is based on PD (polyphase downsampling) multiple description coding. The only changes to H.264 are inserting PD before the DCT process and having new data partitioning NAL units. A coded slice is sent on 3 data partitioning NAL units. A header NAL unit contains motion vectors and block modes. Each of the other two NAL units contains a description generated by PD multiple description coding. The experimental results on all 9 of the test sequences of JVT SVC show that the proposed method gives 0.5 to 5 dB enhancement over the existing H.264 FMO checker board mode with motion vector based error-concealment.

A most probable scan mode decision for H.264/AVC inter picture coding

Video coding standard H.264/AVC employs zig-zag scan and field scan to map transform coefficients from two dimensional matrix to one dimensional array. The zig-zag scan is used for transform coefficients in frame macroblocks (MBs), and the field scan is used in field MBs. This paper presents a most probable scan mode (MPSM) decision for H.264 inter picture coding. Besides the two scans, a horizontal scan is illustrated. For a 4×4 block, one out of the three scan modes is selected as the most probable one, and is used for the transform coefficient scan. Furthermore, an improved MPSM decision is discussed. Simulation results report that the proposed method yields an average of 1.15% bit rate reduction over the H.264 baseline profile.

Macroblock-Based Adaptive Loop Filter for Video Compression

In this paper, we propose an adaptive loop filter that can work on macroblock-level encoding. First, we calculate the filter coefficients that minimize the mean square error between original and encoded frames, and we then apply 2D-filtering to the encoded macroblock using the coefficients. In addition to alleviating blockness, the proposed filtering process improves coding efficiency by allowing the use of filtered pixels as reference pixels for consecutive macroblock coding. The experimental results show that the proposed macroblock-based adaptive loop filter (MBALF) methods can achieve 6% bitrate reduction on average, as compared with the H.264/AVC high profile.

Polyphase Downsampling Based Redundant Picture Coding for SVC Error Resiliency

Scalable video coding (SVC) is currently being developed as an extension to the H.264/AVC by the Joint Video Team (JVT). The SVC addresses coding schemes for reliable video delivery over heterogenous network for diverse clients using available system resources. SVC error resilience strives for reliability as well as efficiency of video transmission under unreliable network conditions. This paper presents a polyphase downsampling based redundant picture coding for SVC error resilience. Simulation results report that the proposed method outperforms the SVC error concealment method by 1.7 dB on average in terms of the PSNR.