Video Compression for Surveillance Application using Deep Neural Network

We present a new approach to video compression for video surveillance by refining the shortcomings of conventional approach and substitute each traditional component with their neural network counterpart. Our proposed work consists of motion estimation, compression and compensation and residue compression, learned end-to-end to minimize the rate-distortion trade off. The whole model is jointly optimized using a single loss function. Our work is based on a standard method to exploit the spatio-temporal redundancy in video frames to reduce the bit rate along with the minimization of distortions in decoded frames. We implement a neural network version of conventional video compression approach and encode the redundant frames with lower number of bits. Although, our approach is more concerned toward surveillance, it can be extended easily to general purpose videos too. Experiments show that our technique is efficient and outperforms standard MPEG encoding at comparable bitrates while preserving the visual quality.

On SPARC LEON-2 ISA Extensions Experiments for MPEG Encoding Acceleration

This paper presents the necessary steps to modify the implementation of the SPARCV8 architecture to enhance it with multimedia-oriented instructions. The purpose is improving video compression performance without designing dedicated coprocessors. We investigate the complexity of modifying a standard processor instruction set and show that, although not trivial, this is feasible in a few weeks. We implemented 12 new instructions and use some of them to optimize the computation of a demanding step of the MPEG encoding. The result is a performance increase of 67% in the execution of a part of this algorithm, allowing us to expect a 30% speedup in the execution of an MPEG video compression. The area increase of the integer unit is about 18% and the clock frequency is not significantly modified in an LEON-2 implementing 6 among 12 of the new instructions.

PERCEPTUALLY ADAPTIVE VIDEO WATERMARKING USING MOTION ESTIMATION

Video watermarks must maintain picture quality and withstand image processing, and resolving the tradeoff between these conflicting requirements has been one of the central problems in research on video watermarking. This paper shows that watermarks are less perceptible where picture contents are moving and proposes a new criterion for measuring watermark imperceptibility from motion vectors and deformation quantities. It also proposes a watermarking method that uses this criterion to allocate watermarks to picture areas adaptively. Experimental evaluation showed that the proposed method reduces degradation in picture quality and that, for the same picture quality after MPEG encoding and decoding, between 30% and 40% more watermarks can be embedded, improving detection reliability after MPEG processing by an average of 9.5%. The proposed method would be most effective when used with MPEG encoders because they already have the necessary motion estimation functions.