1-7hit |
Zheng FANG Tieyong CAO Jibin YANG Meng SUN
Saliency detection is widely used in many vision tasks like image retrieval, compression and person re-identification. The deep-learning methods have got great results but most of them focused more on the performance ignored the efficiency of models, which were hard to transplant into other applications. So how to design a efficient model has became the main problem. In this letter, we propose parallel feature network, a saliency model which is built on convolution neural network (CNN) by a parallel method. Parallel dilation blocks are first used to extract features from different layers of CNN, then a parallel upsampling structure is adopted to upsample feature maps. Finally saliency maps are obtained by fusing summations and concatenations of feature maps. Our final model built on VGG-16 is much smaller and faster than existing saliency models and also achieves state-of-the-art performance.
Heon HUH Feng LU James V. KROGMEIER
In OFDM systems, link performance depends heavily on the estimation of symbol-timing and frequency offsets. Performance sensitivity to these estimates is a major drawback of OFDM systems. Timing errors destroy the orthogonality of OFDM signals and lead to inter-symbol interference (ISI) and inter-carrier interference (ICI). The interference due to timing errors can be exploited as a metric for symbol-timing synchronization. In this paper, we propose a novel method to extract interference components using a DFT of the upsampled OFDM signals. Mathematical analysis and formulation are given for the dependence of interference on timing errors. From a numerical analysis, the proposed interference estimation shows robustness against channel dispersion.
Takuro YAMAGUCHI Masaaki IKEHARA
Image interpolation is one of the image upsampling technologies from a single input image. This technology obtains high resolution images by fitting functions or models. Although image interpolation methods are faster than other upsampling technologies, they tend to cause jaggies and blurs in edge and texture regions. Multi-surface Fitting is one of the image upsampling techniques from multiple input images. This algorithm utilizes multiple local functions and the weighted means of the estimations in each local function. Multi-surface Fitting obtains high quality upsampled images. However, its quality depends on the number of input images. Therefore, this method is used in only limited situations. In this paper, we propose an image interpolation method with both high quality and a low computational cost which can be used in many situations. We adapt the idea of Multi-surface Fitting for the image upsampling problems from a single input image. We also utilize local functions to reduce blurs. To improve the reliability of each local function, we introduce new weights in the estimation of the local functions. Besides, we improve the weights for weighted means to estimate a target pixel. Moreover, we utilize convolutions with small filters instead of the calculation of each local function in order to reduce the computational cost. Experimental results show our method obtains high quality output images without jaggies and blurs in short computational time.
Atsushi KOBAYASHI Shingo KASHIMA Toshihiko KATO
This paper presents a method of measuring the traffic of IP multicast streaming, such as IPTV, by using IPFIX and PSAMP. The IP multicast streaming service has recently become one of the popular network services, but no IP multicast operation method has been established yet. In particular, traffic measurement of IP multicast streaming encounters two challenges: as monitoring a multicast path tree and QoS measurement. There is no method to monitor them continuously in large-scale networks. Hence, we explore measurement structure suitable for large-scale networks, and then using IPFIX and PSAMP, we propose an efficient a network-level quality measurement method for IPTV and a method to extract multicast path tree data. In addition, we demonstrate the feasibility of the measurement method by evaluating a prototype system.
Atsushi KOBAYASHI Keisuke ISHIBASHI
We present the development of a VoIP quality of service (QoS) measurement system that enables operators to diagnose a QoS degradation segment. Our system uses a flow-based passive measurement method to fulfill the requirement for QoS measurement in large-scale IP networks. In particular, we adopt an access control list (ACL)-based filtering function that selects traffic to monitor and develop a function for correlating signals and media data records. This correlation function is required to dynamically configure ACL-based filtering for monitoring media streams whose port numbers are determined by a signaling protocol. To improve the scalability of existing measurement systems, we also develop a hardware-based filtering engine on a commercial switch as well as a mediation box that performs QoS calculation based on traffic records exported by the engine in a distributed manner. We demonstrate the feasibility of the measurement system by evaluating a prototype system.
Do Nyeon KIM Yoonsik CHOE K.R. RAO
Fast schemes for compressed-domain image size change, are proposed. Fast Winograd DCTs are applied to resizing images by a factor of two to one. First, we speed up the DCT domain downsampling scheme which uses the bilinear interpolation. Then, we speed up other image resizing schemes which use DCT lowpass truncated approximations. The schemes proposed here reduce the computational complexities significantly, while there is no difference in the overall quality of the images compared to previous works.
Satoshi UEMURA Miki HASEYAMA Hideo KITAJIMA
This letter presents a significant property of the mapping parameters that play a central role to represent a given signal in Fractal Interpolation Functions (FIF). Thanks to our theoretical analysis, it is derived that the mapping parameters required to represent a given signal are also applicable to represent the upsampled signal of a given one. Furthermore, the upsampled signal obtained by using the property represents the self-affine property more distinctly than the given signal. Experiments show the validity and usefulness of the significant property.