As an important type of image classification task, fine-grained image classification is of great practical interest. The major challenges of fine-grained image classification come from (1) The available training datasets are lack of object localization annotations which are highly time-consuming (2) The object discriminating a class generally only takes a very small portion of a whole image. In this paper, a spatial saliency extraction (SSE) approach is proposed to crop the possible attention areas inside a whole image to improve the fine-grained image classification performance. The proposed algorithm has been applied to the recently released Chest-X-ray14 dataset [1], which is highly imbalanced and weakly annotated. The SSE method offers the state-of-the-art classification performance.

A novel parity-check matrix design for low-density parity-check (LDPC) codes is described. By eliminating the routing problem associated with LDPC codes, the design results in a small implementation area, and the codes have outstanding error-rate performance close to the Shannon limit for a wide range of code rates, from 1/4 to 9/10, and for various modulation schemes such as binary phase-shift keying (PSK), quaternary PSK, 8-PSK, 16-amplitude PSK (APSK), and 32-APSK. As a result, LDPC codes designed with this method have been standardized for next-generation digital video broadcasting.

Tremendous progress has been made on channel coding techniques during the last decade, starting from the invention of turbo codes and the rediscovery of the low-density parity check (LDPC) codes. These codes essentially closed the gap between the Shannon capacity limit and practical implementation. Commercial adoption of these new coding techniques, however, depends heavily on the introduction of the technology into communication standards. In this paper, we briefly review the considerations when turbo codes and LDPC codes were introduced to the 3/sup rd/ generation wireless standards (3GPP and 3GPP2), and the digital video broadcast for satellite standard (DVB-S2), respectively. We then examine a few possible approaches to take advantage of these powerful codes for future wireless channels.

The performance of linear minimum mean-square error (MMSE) multiuser receivers in a dual-rate synchronous direct-sequence code-division multiple-access (DS-CDMA) system is investigated using the random spreading sequence analysis. Multicode (MC) systems and different variants of variable spreading length (VSL) systems are studied. User capacity regions are obtained for these systems.

Motivated by the observation that in some applications, it is highly desirable for the physical layer itself to have a user-identification feature instead of obtaining user identity from the media access control layer, we propose a method to give a convolutionally or turbo-coded system an identity. The user-identification algorithm embeds a periodic identification sequence into the convolutionally coded data sequence in the transmitter, and detects and removes the identification sequence with the assistance from the forward error-correction decoder in the receiver. It does not consume extra transmission bandwidth by tailoring the identification sequences to the codes.

This paper addresses data-aided (DA) synchronization, in which the reference parameter acquisition is aided by a training sequence known to the receiver. The Cramer-Rao lower bound (CRB) for the DA timing and/or carrier phase recovery is presented. For DA parameter estimation, the CRB typically varies with the training sequence. This indicates that different training sequences offer fundamentally different performance. In the literature, the widely cited closed-form CRB for timing and carrier phase recovery was derived under the assumption that the training sequence is independent and identically distributed (i.i.d.) and sufficiently long. We derive a closed-form CRB for timing and carrier phase recovery with respect to an arbitrary training sequence and pulse shaping function for the over and under sampling cases. It turns out that the CRB is a weighted summation of the aperiodic correlation of the training sequence and the weighting factor is determined by the pulse shaping filter. Therefore, this paper reveals the fundamental link between a training sequence and its corresponding performance limit.

Many issues in signal processing involve the inverses of Toeplitz matrices. One widely used technique is to replace Toeplitz matrices with their associated circulant matrices, based on the well-known fact that Toeplitz matrices asymptotically converge to their associated circulant matrices in the weak sense. This often leads to considerable simplification. However, it is well known that such a weak convergence cannot be strengthened into strong convergence. It is this fact that severely limits the usefulness of the close relation between Toeplitz matrices and circulant matrices. Observing that communication receiver design often needs to seek optimality in regard to a data sequence transmitted within finite duration, we define the finite-term strong convergence regarding two families of matrices. We present a condition under which the inverses of a Toeplitz matrix converges in the strong sense to a circulant matrix for finite-term quadratic forms. This builds a critical link in the application of the convergence theorems for the inverses of Toeplitz matrices since the weak convergence generally finds its usefulness in issues associated with minimum mean squared error and the finite-term strong convergence is useful in issues associated with the maximum-likelihood or maximum a posteriori principles.

Watermarking is a process that embeds one signal into another signal to provide a hidden identity. It is widely used in multimedia signal processing for copyright protection purpose. Motivated by the observation that in some applications it is highly desirable for the physical layer itself to have some user identifying features instead of obtaining such information from the media access control (MAC) layer, we extend the watermarking idea and propose a method to give a convolutionally or turbo coded system a hidden identity. The proposed algorithm embeds a periodic watermark sequence into the convolutionally coded data sequence in the transmitter and detects the watermark using the channel error information supplied by the forward error control (FEC) module in the receiver. It does not consume any extra transmission bandwidth by properly designing the watermark sequence using the properties of convolutional codes. The algorithm has several potential applications in wireless and satellite communications. It is easy to implement and fits generic modems.

This paper addresses the channel estimation problem for slow frequency-selective fading channels using training sequences and the maximum likelihood (ML) approach. In the literature people usually assume a symbol period spaced delay-tapped-line model and additive white Gaussian noise (AWGN). Due to the pre-filtering in the receiver front end, if the sampling rate is larger than one sample per symbol or sampling epoch is unknown (i.e., the timing information is unavailable), the AWGN model is not valid anymore. A more general ML channel estimation method using the discrete Fourier transform (DFT) is derived for colored Gaussian noise and over-sampling. A similar idea can be adopted to derive the ML joint carrier phase and timing offsets estimation algorithm.

This paper concerns the Cramer-Rao lower bound (CRB) for the data-aided (DA) timing and/or phase recovery, i.e., the synchronization parameter acquisition is aided by a training sequence known to the receiver. For the DA parameter estimation, the CRB typically varies with the training sequence. This indicates that different training sequences offer fundamental different performance. We derive a closed-form formula of the CRB for timing and phase recovery with respect to any particular training sequence. The bound illustrates the close relation between the training sequence and the fundamental limit on timing and phase synchronization. It provides additional insights on the training sequence design.

Wireless data applications have just begun to gain popularity in the marketplace. The rollout of 3rd generation (3G) wireless networks is anticipated in the next few years. We examine wireless data networks that are complementary to the 3G wireless networks, their technologies and the challenges. In particular, we present some of the industry initiatives in the millimeter wave frequency bands for broadband and satellite delivered fixed wireless data network. Furthermore, we offer our perspectives on how these different networks will evolve and inter-work together to provide a seamless experience for the end users. Fixed data networks need to be competitive with wire-line services, this requires us to push advanced signal processing technologies further. We present the latest advances that we have made in channel coding and signal processing technologies to approach that goal.

The telecommunication industry is evolving at an unprecedented pace. As the PCS operators are just rolling out their network, the ITU (International Telecommunication Union) has already initiated the development of the third generation mobile communication standards. This paper presents an overview of the design issues and technical innovations as being developed during the process of the standard developments.

We consider the construction and analysis of linear block codes for M-ary phase-shift keying that can be decoded without carrier phase synchronization. Under these circumstances, the function that has a significant impact on performance is the noncoherent distance, analogous to the Euclidean distance for the coherent case. The major difficulty in constructing and analyzing such codes lies in the fact that the noncoherent distance is not a true metric. For this reason, prior work mainly relies on numerical approaches to search for good codes and to determine the corresponding minimum noncoherent distance. We first present a theorem that links the noncoherent distance with the Euclidean and Lee (1958) distances. This theorem allows us to construct good codes and determine their minimum noncoherent distances analytically. Based on this theorem, we classify the codes whose duals consist of a cyclic group. These codes are of minimum redundancy. We further investigate codes with the flavor of Hamming and shortened Hamming codes. Many of these new codes provide significantly larger coding gains than previously known codes. Linear codes derived from code-division multiple-access (CDMA) sequences are considered as well. These codes in general provide rather large coding gains. Finally, an algorithm is introduced that can be appended to any suboptimal decoding technique to enhance the performance.

Signature sequences with good even even and odd (or polyphase) correlations are crucial for asynchronous code-division multiple access (CDMA). When the data sequence is random, the even and odd (or polyphase) correlations are equally important. However, for most known signature sequences, only their even correlations were analyzed. It appears that determining the odd (or the polyphase) correlations is generally a very hard problem since the odd (or the polyphase) correlations depend on the phases of the signature sequences. Sole (1989), Boztas, Hammons, and Kumar (1992) found a family of quadriphase sequences that are asymptotically optimal. These sequences gain a factor /spl radic/2 in terms of their maximum periodic even correlations when compared with the best possible binary phase-shift keying (BPSK) sequences. We find the optimal phases of these sequences. The optimality is in the sense that at these phases, the mean square values of the even, odd, and the polyphase correlations are minimal, and achieve the Welch (1974) bound-equality simultaneously. Furthermore, we show that at these phases, the average user interference of these sequences is always smaller than that of the ideal random signature sequences. Comprehensive analytical and numerical results show that good phase sequences can offer a nonnegligible amount of gain over bad phase sequences at modest and high signal-to-noise ratios.

It is known that both Viterbi and sequential decoding of convolutional codes can be greatly simplified by employing punctured convolutional codes, which are obtained by periodically deleting a part of the bits of a low-rate convolutional code. Even if the original low-rate convolutional code is noncatastrophic, some deleting maps may result in rate (n-1)/n catastrophic punctured encoders. An algorithm is presented to identify such encoders when the original rate 1/b encoder is antipodal. The major part of the algorithm solves a linear equation of /spl nu/+1 variables, where /spl nu/ is the constraint length of the original rate 1/b code.

New multilevel constructions of the Golay code and the Leech lattice are presented. These are derived from the Turyn construction and the "holy construction" with the octacode as the glue code. Further, we show that the "holy construction" of the Leech lattice with the octacode as the glue code is essentially different from the permuted Turyn construction, although both constructions rely on the octacode. The Turyn construction is based on an "odd" type of the octacode, whereas any type of the octacode can be used in the "holy construction." Moreover, the multilevel representation of the "holy construction" leads to a novel lattice partition chain. Based on these structures, we derive new bounded-distance decoders for the Golay code and the Leech lattice whose effective error coefficient is smaller than that of any previously known bounded-distance decoder. We provide a general theorem for computing the effective error coefficient of coset decoding with bounded distance decoding for the subcode.>

Multilevel constructions of the binary Golay code and the Leech lattice are described. Both constructions are based upon the projection of the Golay code and the Leech lattice onto the (6,3,4) hexacode over GF(4). However, unlike the previously reported constructions, the new multilevel constructions make the three levels independent by way of using a different set of coset representatives for one of the quaternary coordinates. Based upon the multilevel structure of the Golay code and the Leech lattice, efficient bounded-distance decoding algorithms are devised. The bounded-distance decoder for the binary Golay code requires at most 431 operations. As compared to 651 operations for the best known maximum-likelihood decoder. Efficient bounded-distance decoding of the Leech lattice is achieved by means of partitioning it into four cosets of Q/sub 24/, beyond the conventional partition into two H/sub 24/ cosets. The complexity of the resulting decoder is only 953 real operations on the average and 1007 operations in the worst case, as compared to about 3600 operations for the best known in maximum-likelihood decoder. It is shown that the proposed algorithms decode correctly at least up to the guaranteed error-correction radius of the maximum-likelihood decoder. Thus, the loss in coding-gain is due primarily to an increase in the effective error-coefficient, which is calculated exactly for both algorithms. Furthermore, the performance of the Leech lattice decoder on the AWGN channel is evaluated experimentally by means of a comprehensive computer simulation. The results show a loss in coding-gain of less than 0.1 dB relative to the maximum-likelihood decoder for BER ranging from 10/sup -1/ to 10/sup -7/.>

For the additive white Gaussian noise channel, there is a gap between the channel capacity and the highest achievable rate of equiprobable uniformly spaced one-dimensional signaling. It is commonly believed that approaching channel capacity requires the constituent one-dimensional signal points to have a Gaussian probability distribution. It is shown that the channel capacity can be achieved by equiprobable signaling with geometrical, Gaussian-like signal sets. Construction of these signal constellations is explicitly given. This result implicates that it is possible to approach channel capacity without using any kind of shaping codes.>