Decoder driven side information generation using ensemble of MLP networks for distributed video coding

Abstract

This paper proposes an ensemble of multi-layer perceptron (MLP) networks for side information (SI) generation in distributed video coding (DVC). In the proposed scheme, both three-layer and four-layer MLP structures are used to form the ensemble model. The proposed model includes four sub-modules. The first sub-module involves the training of the individual networks. The second sub-module selects ‘M’ number of trained MLPs based on the mean square error (MSE) performance metric. Next, the third sub-module involves the testing phase of each of the selected MLPs. Finally, in the last sub-module, the overall ensemble SI is generated using a dynamically averaging (DA) method. The primary goal of this work is to minimize the estimation error between the SI and the corresponding Wyner-Ziv (WZ) frame so that the overall efficiency of DVC codec can be increased. The proposed scheme is evaluated with respect to different parameters such as Rate-Distortion (RD), Peak Signal to Noise Ratio (PSNR), Structural Similarity Index (SSIM), and number of parity requests made per estimated frame. The evaluation indicates that the proposed ensemble model shows better generalization capabilities with improved PSNR (in dB) as compared to each of the individual selected networks. Additionally, the comparative analysis also exhibits that the proposed SI generation scheme generates better SI frames in comparison with the contemporary techniques. Further, using a statistical test, namely, ANOVA with significance level of 5%, it has been validated that the proposed technique yields a significant enhancement in the performance as compared to that of the benchmark schemes.