Scalable FPGA Hardware Acceleration for H.264 Motion Estimation

The H.264 video compression standard uses enhanced Motion Estimation (ME) features to improve both the compression ratio and the quality of compressed video. The two primary enhancements are the use of Variable Block Size Motion Estimation (VBSME) and multiple reference frames. These two additions greatly increase the computational complexity of the ME algorithm, to the point where a software based real-time (30 frames per second (fps)) implementation is not possible on present microprocessors. Thus hardware acceleration of the H.264 ME algorithm is necessary in order to achieve real-time performance for the implementation of the VBSME and multiple reference frames features. This thesis presents a scalable FPGA-based ME architecture that supports real-time H.264 ME for a wide range of video resolutions ─ from 640x480 VGA to 1920x1088 High Definition (HD). The architecture contains innovations in both the data-path design and memory organization to achieve scalability and real-time performance on FPGAs. At 37% FPGA device utilization, the architecture is able to achieve 31 fps performance for encoding full 1920x1088 progressive HDTV video.

Scalable FPGA Hardware Acceleration for H.264 Motion Estimation

The H.264 video compression standard uses enhanced Motion Estimation (ME) features to improve both the compression ratio and the quality of compressed video. The two primary enhancements are the use of Variable Block Size Motion Estimation (VBSME) and multiple reference frames. These two additions greatly increase the computational complexity of the ME algorithm, to the point where a software based real-time (30 frames per second (fps)) implementation is not possible on present microprocessors. Thus hardware acceleration of the H.264 ME algorithm is necessary in order to achieve real-time performance for the implementation of the VBSME and multiple reference frames features. This thesis presents a scalable FPGA-based ME architecture that supports real-time H.264 ME for a wide range of video resolutions ─ from 640x480 VGA to 1920x1088 High Definition (HD). The architecture contains innovations in both the data-path design and memory organization to achieve scalability and real-time performance on FPGAs. At 37% FPGA device utilization, the architecture is able to achieve 31 fps performance for encoding full 1920x1088 progressive HDTV video.

Temporal dynamics of ensemble statistics calculation using a neural network model

Computing summary or ensemble statistics of a visual scene is often automatic and a hard necessity for stable perceptual life of a cognitive agent. Although computationally the process should be as simple as applying a filter as it were to a perceived scene, the issue of mechanism of summary statistics is complicated by the fact that we can seamlessly switch from summarizing to individuation while computing the ensemble averages across multiple reference frames. In the current work we have investigated the possibility of a neural network that can also switch between individuation and summarization. We have chosen a computational model previously used for enumeration/individuation (Sengupta et al, 2014) in order to show possibility of extracting summary statistics using two different measures from the network. The results also shed a light on possible temporal dynamics of ensemble perception.