Published Paper Details:

The consumer demand for video applications is rapidly increasing in the present digital market. This necessitates the compression of the video streams and the HEVC (High Efficiency Video Coding) standard significantly reduces network traffic and bandwidth needs. The coarse quantization that results in many artifacts is the main cause of the compressed video's quality loss. The goal of this work is to improve the quality of HEVC compressed decoded video using an Improved Attention U-Net architecture and to detect moving objects in the improved HEVC decoded video using a modified Mask R-CNN object detection approach. Deep learning based Denoising model can be used as a preprocessing technique in prior to Improved Attention U-Net GAN architecture. In the improved video, moving objects are found using the Mask R-CNN model. Using a post-processing technique like Non-Maximum Suppression (NMS), duplicate detections can be eliminated and the model's accuracy can be increased. The proposed method of quality enhancement demonstrated a maximum improvement of 7.71 dB for PSNR and a maximum improvement of 6.01% for SSIM, when compared to SRGAN (Super Resolution Generative Adversarial Network), MLDCNN (Multi Layered Deep Convolutional Neural Network) and Deep Learning with Super Interpolation - Laplacian Filter (DLSI-LF) methods. In comparison to Faster R-CNN and Single Shot multibox Detection (SSD) approaches, the proposed object detection method achieves a maximum precision of 0.9812, a maximum recall of 0.9789, a maximum F-measure of 0.9800 and a maximum accuracy of 0.9984 with a maximum improvement of 8.84%.