Optimized quantization for image subband coding

This paper presents a complete coding system for compression of natural color still images represented with 24 bits per pixel. The original 8-bit red, green, and blue (RGB) image components are rst transformed to a luminance (Y) and two chrominance components (Cb, Cr) according to the ITU-R Recommendation 601 1]. Each of the Y, Cb, and Cr tristimulus values are then decomposed separately in a subband lter bank with di erent lters for the luminance and the chrominance components. The Y-component requires a signi cantly higher bit rate than the Cb-and Cr-components, which can be downsampled by a factor of 10 to 20. For high quality image compression the downsampling factor must not be too large to avoid color artifacts. The luminance and chrominance components are quantized in uniform threshold quantizers with step size adapted to the signal statistics and the human visual system. A bank of entropy coders adapted to the components' subband signal statistics is used for bit e cient representation of the quantized samples. The coding results shows that a transparent image quality can be obtained at about 1.2 bits per pixel for the Mandrill image.

Journal of Visual Communication and Image Representation, 2017

Uniform quantization schemas with dead zone are widely used in image and video codecs. The design of these quantizers affects to the final R/D performance, being two of the quantizer parameters the responsible for that variations: the dead zone size and the reconstruction point location inside each quantization step. In this work we tune the quantizer to obtain the optimum quantization parameters that provide the best R/D behavior for different quality metrics and rate ranges. Based on a representative image set, we provide the quantization parameters to encode general imagery, with a R/D performance close to the optimum one. The same study was done including the Contrast Sensitivity Function in the quantization stage. After an exhaustive experimental test, the results show that the estimated quantization parameters are able to provide bit rate savings up to 11% at low and moderate bit rates without additional computational cost.

1994

Digital images are replacing analogue images such as photographs and x-rays in many different fields. Compression of these digital images is desirable for efficient storage and transmission. Subband coding has proved an effective method of image compression. This thesis investigates subband analysis structures and filters which are optimised for still image compression. A m o n g other results it is shown that a high coding gain, based on a typical image model, and good spatial localisation are desirable filter bank characteristics for subband image coding. Assuming a high bit rate it is well known that the Karhunen-Loeve Uansform (KLT) is the optimum orthogonal block Uansform in terms of a coding gain metric. It is shown further that the K L T is the optimum invertible block Uansform using the unified coding gain. The coding gain metric is examined under a rate constraint. It is shown that for highly correlated sources increased low frequency subband resolution is required for optimum performance at low rates as compared to high rates: a result that is corroborated using a practical subband coder. Subband filters (CQF's) that globally maximise the coding gain for all two-band pe reconstruction orthogonal filter banks are derived. Various characteristics of these filters are predicted using a new theorem on the zeros of an eigenvector of a symmetric Toeplitz matrix corresponding to the minimum (maximum) eigenvalue. These filters are shown to enjoy the three properties of the K L T : namely m a x i m u m coding gain, minimum basis restriction error, and subband decorrelation. It is also shown that there is some freedom to select different impulse responses. The design of m a x i m u m gain filters is extended to include filters consUained to certain subspaces. For example m a x i m u m gain wavelets may be designed. A modified two-climensional discrete wavelet Uansform (DWT) is proposed based on a typical image model. A generic subband quantisation and encoding method suitable for any subband structure is inuoduced. This method is essentially a generalisation of the J P E G quantisation and encoding method and has good spatial adaptation properties. Using the generic subband quantisation and encoding method various subband analysis structures and filters are compared for still image compression. The best orthogonal iii filters, Daubechies wavelets a n d m a x i m u m gain filters designed using an i m a g e source model, and the discrete cosine transform ( D C T ) perform in a similar manner. The filters with the minimum spatial (time) width perform better than other impulse responses in a mean square error sense and exhibit significantly less ringing. The performance of cosine modulated filter banks, with poorer spatial resolution, is slightly inferior to the D C T . Preliminary investigations show that biorthogonal filters, with a smaller spatial width and higher coding gain, can outperform the best orthogonal filters, especially at low rates. These biorthogonal filters also exhibit minimal ringing. Finally, the modified D W T is shown to be superior to the D W T for head and shoulders type images.

Journal of Visual Communication and Image Representation, 1998

adapt the conventional image coding schemes to the visual This paper describes a new methodology in the design of system, or we can try to design new compression algopsychovisual quantizers for a given visual subband image codrithms. ing scheme. The quantizer design needs to specify a set of In the first case, the use of visual properties, such as the decision thresholds (T k , k ‫؍‬ 0,. .. , l) and reconstruction levels visibility thresholds in the quantization of predefined image (L k , k ‫؍‬ 0,. .. , l ؊ 1) for which any input level lying between representations, involves a lot of difficulties and leads usuthe thresholds T k and T k؉1 is mapped to L k , the kth reconstrucally to a certain degree of approximation. As an example tion level. To specify the optimal quantization, psychophysical we quote the optimized quantizer designed by Vanderdope experiments based on the visibility of quantization noise have [1] for image subband coding. The quantization steps for been conducted. The contrast masking function has been detereach subband were defined as inversely proportional to mined by considering complex signals, spatial frequency aspects the square root of the decimation factor and a weighting of the quantized signals, and the spatial combination of degradations. The stimuli used have been characterized by their factor. The latter depends on parameters such as the dislocal band-limited contrast. With this representation, within a tance of the observer from the screen, the sensitivity funcsubband, the first important result is that the decision threshtion, the equivalent synthesis filter and the sampling paolds and the reconstruction levels of the quantization follow a rameters of the picture. On the one hand, however, in the linear law regardless of the importance of the contrast. From definition of this weighting factor it has been assumed that one subband to another the quantization step varies according the quantization noise is not correlated with the values to to radial spatial frequency and orientation. In order to examine be quantized. On the other hand the quantization errors the masking effects between subbands, further experiments of successive samples in the same subband are considered have been conducted and two other important results have to be uncorrelated. This hypothesis modifies the noise been obtained. First, when it occurs, masking does not modify structure and leads to limited performances. the linear behavior of the optimal quantization law. Secondly, masking is due to the adjacent angular subbands rather than Conversely, in the second approach one tries to construct the adjacent radial ones, even if a slight increase in the quantizaan image representation which fits, as much as possible, tion step is observed in this last case. the one used in the human visual system and then one Key Words: psychovisual quantizer, visual subband; maskattempt to quantize this representation with the precision ing; local band-limited contrast; visibility of quantization required by the desired image quality. An interesting exnoise.

This paper presents a subband coder for compression of color still images. The red, green, and blue (RGB) color components are rst transformed to a luminance-chrominance representation, and then each of the three image components are decomposed by an FIR analysis lter bank. The two chrominance components are decimated by a factor of 10:1 or 20:1, depending on the bit rate. The coder is based on dynamic classi cation of blocks of subband samples into 5 classes according to the blockwise mean square value. Each subband sample is quantized in a uniform threshold quantizer, and the quantizer index coded in an arithmetic coder matched to the statistics of the corresponding class. The dynamic classi cation scheme gives a signi cant rate-distortion gain as compared to a scheme without classi cation using the same lter banks and coding methods.