研究目的
Investigating an efficient LightField (LF) image coding scheme based on Convolutional Neural Networks (CNN) and Linear Approximation (LA) to reduce the large amount of data produced by LF imaging systems and improve compression efficiency.
研究成果
The proposed coding solution based on CNN and linear approximation for 4D-LF images significantly reduces the bit-rate by 51% compared to the HM encoder and increases the visual quality of the novel views. It also achieves a bitrate reduction of 30% on average compared to the state-of-the-art LF image compression solution (LA-32). Future research directions include enhancing the coding performance with more advanced DL systems and evaluating the complexity of the proposed solution at both encoder and decoder sides.
研究不足
The proposed solution shows inconsistent performance for certain LF images, possibly due to the interpolation interval of the BD-BR metric. Future work includes enhancing the DL systems and CNN parameters with reinforcement learning and investigating the optimal coding of residuals at the encoder side.
1：Experimental Design and Method Selection:
The proposed scheme first encodes a sparse set of views using the latest hybrid video encoder (JEM) and estimates a second sparse set of views using linear approximation. At the decoder side, a Deep Learning (DL) approach is used to estimate the whole LF image from the reconstructed sparse sets of views.
2：Sample Selection and Data Sources:
Nine LF images from Ecole Polytechnique Federale de Lausanne (EPFL) LF images dataset composed of 8x8 views were selected for the experiment.
3：List of Experimental Equipment and Materials:
The experiment utilized the JEM video codec, HEVC reference Model (HM), and a trained CNN block for view synthesis.
4：Experimental Procedures and Operational Workflow:
The scheme involves encoding a sparse set of views, estimating another set using linear approximation, and synthesizing the remaining views using a trained CNN. The performance was assessed using BD-BR and WPSNR metrics.
5：Data Analysis Methods:
The BD-BR and WPSNR metrics were used to evaluate the coding performance and visual quality of the reconstructed LF images.
独家科研数据包，助您复现前沿成果，加速创新突破
获取完整内容
