A dictionary generation scheme for block-based compressed video sensing

Abstract

Compressed sensing is a novel technology that exploits sparsity of a signal in a transform domain to perform sampling below the Nyquist rate, and has great potential in video coding applications for its low-complexity. However, the traditional orthonormal basis cannot be adopted to provide a sparse enough representation for compressed video sensing. Therefore, how to use the temporal/spatial redundancy in video is the main challenge. In this paper, we propose a dictionary generation scheme for block-based compressed video sensing. By means of motion estimation in measurement domain, the dictionary is initialized using blocks extracted from the reference frame as the atoms. Then an estimation of the current frame can be obtained, which is in turn employed to update the dictionary. The proposed algorithm provides a more accurate dictionary for the sparse representation of video in an iterative fashion. And the experimental results show that our proposal offers comparable performance to other existing methods, with a 0.8dB to 2dB improvement in the average PSNR.

Distributed Compressed Video Sensing with Joint Optimization of Dictionary Learning and <i>l</i>₁-Analysis Based Reconstruction

“…However, from the perspective of distributed CS [27], the performance of joint reconstruction can be achieved better by using smaller MR NK and larger MR K , if the temporal redundancy between frames could be exploited at the decoder. Consequently, in the DCVS framework, key frames can be undersampled at an increased MR K in relation to MR NK [8], [11], [17]. Thus, we use the similar conditions described in [15], [16] in our experiment, taking the following two cases MR K = MR NK and MR K > MR NK into consideration.…”

“…In [16], the challenges involved in the transmission of CS-based video over wireless multimedia sensor networks were discussed by Pudlewski et al, and a cross-layer system that jointly controls the video encoding rate, the transmission rate. Additionally, a method to generate dictionary for video sampling based on CS was proposed in our previous work [17], an adaptive alternating direction method of multipliers (ADMM) with its application to compressed video sensing was presented in [18], [19], and more recently, a joint sampling rate and bit-depth optimization framework was proposed in [20]. Nevertheless, how to use the redundant information of temporal and spatial relations by means of side information (SI) at the decoding end to get efficient sparse dictionary, and signal recovery from limited numbers of measurements, is substantially unexplored.…”

Joint Sampling Rate and Bit-Depth Optimization in Compressive Video Sampling

“…; 3) the mean of motion vector (MV) magnitudes ( ), the average of the standard deviation of MV magnitude in each frame ( ) and the average of the standard deviation of MV directions in each frame ( ) (wherein the MV related parameters are determined through the measurement-domain motion estimation [24]); 4) the normalized , and .…”