Journal of Computing and Information Technology - CIT 16, 2008, 3, 145–156 145 doi:10.2498/cit.1000892 An Efficient Coding Method for Teleconferencing Video and Confocal Microscopic Image Sequences Vinay Arya1, Ankush Mittal1, Amit Pande2 and R. C. Joshi1 1Dept. of Electronics and Computer Engineering, IIT Roorkee, Uttaranchal, India 2Dept. of Electrical and Computer Engineering, Iowa State University, Ames, USA In this paper we propose a three-dimensional vector scalar quantization schemes and its simple de- quantization-based video coding scheme. The algorithm coder structure. This method of compression uses a 3D vector quantization pyramidal codebook-based model with adaptive pyramidal codebook for compres- has gained interest over conventional means of sion. The pyramidal codebook-based model helps in compression (i.e., DPCM and Transform cod- getting high compression in case of modest motion. The ing [1]) because the conventional methods usu- adaptive vector quantization algorithm is used to train the ally employ some form of scalar quantization. codebook for optimal performance with time. Some of the distinguished features of our algorithm are its excel- Scalar quantization is not optimal as the suc- lent performance due to its adaptive behavior to the video cessive samples in a digital signal are usually composition and excellent compression due to codebook correlated or dependent. A vector can be used approach. We also propose an efficient codebook-based to describe almost any type of pattern, such as post-processing technique which enables the vector quan- tizer to possess higher correlation preservation property. the patterns that arise in digital images. Based on the special pattern of the codebook imposed by Different approaches may be considered for post-processing technique, a window-based fast search (WBFS) algorithm is proposed. The WBFS algorithm video compression. Since lossless compres- not only accelerates the vector quantization processing, sion approaches result in small compression ra- but also results in better rate-distortion performance. tios, we have considered the lossy compression The proposed approach can be used for both telecon- techniques. Several studies dealing with vec- ferencing videos and to compress images obtained from tor quantization-based lossy compression of im- confocal laser scanning microscopy (CLSM). The results ages has been proposed and applied to different show that the proposed method gave higher subjective [ ] and objective image quality of reconstructed images at a image and videos 2-8 . better compression ratio and presented more acceptable Some of the main approaches developed are results when applying image processing filters such as edge detection on reconstructed images. The experi- discussed in this section. Localized model-ba- mental results demonstrate that the proposed method sed approach to teleconferencing using three-di- outperforms the teleconferencing compression standards mensional vector quantization [2] method uses H.261 and LBG-based vector quantization techniques. LBG algorithm for creating a codebook. This Keywords: confocal microscopic images, codebook, codebook creating method does not vary with codebook-preprocessing, rate distortion, vector quan- time. Thus this codebook needs to be trained tization, window-based fast search for optimal performance with varying time, and hence varying input vector characteristics. To overcome this shortcoming, we have proposed an adaptive codebook approach. 1. Introduction Besides good R-D tradeoff, a good encoder should also posses an acceptable level of com- Vector quantization has been widely used in im- plexity. Predictive Vector Quantizer [3] and age and video compression because of its supe- Finite State Vector Quantizer [4] alleviate the rior rate distortion performance over traditional computational burden of a vector quantizer with 146 An Efficient Coding Method for Teleconferencing Video and Confocal Microscopic Image Sequences memory, but at the same time they utilize the Briefly, arterial segments are stained with flu- correlation among neighboring vectors. In [5], orescent markers for the cell nucleus or beta- the inter-index correlation is exploited to code adrenergic receptors. Tissues are then slide the indices more efficiently. Nevertheless, the mounted on the stage of either a NORAN3 (nu- easiest way to exploit the inter-vector correla- clear work) or Leica4 (receptor work) CLSM. tion is to use the DPCM technique in the in- Serial optical sections (x, y plane) are collected dex domain, i.e., to precede the entropy cod- at intervals of 1μm down through the axial plane ing of indices with a DPCM stage. Motivated (z-axis) to produce a ’stack’ of optical planes by these considerations, a simple but efficient which can be processed as a 3D volume. The post-processing technique for adaptive code- book is proposed. The post-processing is done processing, analysis and transfer of the result- to reorder the codebook according to the po- ing data volumes are time inefficient. A robust tential of code words. The reordering is done non-lossy or low distortion lossy compression once for all and is applicable to all existing routine would be of great value for biomedical codebook design methods. The post-processed purpose, e.g. in studying vascular structure [10]. codebook possesses a special structure which Furthermore, the emergence of multi photon in turn leads to decreased first-order differential microscopy as a practical laboratory tool now entropy, i.e., increased correlation between suc- enables even greater depth penetration within cessive indices. This is beneficial to subsequent thick biological samples. This, coupled with DPCM and entropy coding, and it improves the studies involving multiple fluorophores im- the performance of the entire coding system. aged over time period, results in the collection Another merit of the proposed post-processing of data sets approaching 1 GB per experiment. techniques is that it imposes a certain structure Without efficient compression algorithms, man- to the codebook. This structure can be utilized agement of such enormous amounts of data will to accelerate the searching process. Thus it re- be very difficult. lieves the computational requirement of the en- coder. Three-dimensional image compression meth- For microscopic medical images, Burt [6] and ods outperform their two-dimensional counter- Adelson [7] introduced quantization into the im- parts in the sense of higher rate distortion perfor- age pyramid structure and proposed a multi- mance for compressing volumetric image data. resolution lossy compression technique using The state-of-the-art transform-based 3D com- a scalar quantizer. In 2003, Cockshott pro- pressors, such as 3D-SPIHT and 3D-DCT, are posed a vector quantization-based 3D pyramid characterized by their rate control ability, where approach [8]. This method has 3D codebooks the qualities of the image are adjustable with re- × × of size 2 2 4. By using multiple level local- spect to the rates, but are not explicitly control- ized codebooks with variable size, it is possible lable. In [8], the authors discuss the differences to get high compression even if there is some between 3D microscopy raw data and the movie motion in microscopic image sequence. sequence. In microscopy the data corresponds to tracer densities at sub volumes in a 3D grid, Confocal Microscopic Image Sequence with the size of the sub volumes constrained by the microscopy optics. This contrasts with ( ) the data in a movie sequence which has time Confocal laser scanning microscopy CLSM , [ ] which is single photon microscopy, has been varying 3D generating surfaces. In 8 ,theau- available to biomedical scientists for almost 20 thors hypothesize that the higher order statistics years [9]. However, it is only in very recent of 3D microscopy data will differ from those times that affordable computer power has en- of film and that the optimal compression strat- abled biologists to fully exploit the data con- egy will differ from that used in film and video tained within the large (>100Mb) image vol- applications. umes. CLSM is used to collect 3D volumetric data describing the cellular organization and re- In [8], the author presents the use of 3D pyramid ceptor protein distribution through the vascular data structures to compress microscopy data. wall of small segments of human and animal In [11], the author presents vector quantization- arteries. based method in an enhanced image pyramid An Efficient Coding Method for Teleconferencing Video and Confocal Microscopic Image Sequences 147 with error feedback so that the quality of the de- cell has an index. The encoder gives an index compressed image depends only on the encod- to the given vector. On the other end, the de- ing of coefficients from the finest band. This coder generates an output vector from a given achieves some effective reduction of artifacts. index. The task of the encoder is to identify the location of input vector in one of the N geomet- In 3D microscopy, the raw data corresponds to rically specified regions of k spaces. Decoding tracer densities at sub volumes in a 3D grid with is simply a table lookup task fully determined the size of the sub volumes constrained by the by specifying the codebook. A decoder does microscopy optics. The data is typically digi- not need to know the geometry of the partition tized as a sequence of 2D images i.e., a three to perform his job. Vector quantization has been dimensional array with time, x and y axis as the [ ] widely used in image-coding systems because three dimensions 8 . of its superior rate distortion (R-D) performance The paper is organized as follows. The vector over traditional scalar quantization schemes and quantization and adaptive codebook procedures simple decoder structure. are explained in Section 2. 3D vector quantiza- tion used for compression and encoding of tele- conferencing videos and confocal microscopic image sequences is described in Section 3. An efficient codebook post processing technique and a window-based fast search algorithm for vector quantization are explained in Section 4. The experimental results and analysis are given in Section 5.