Nowadays we are facing the transition from desktop PCs to mobile devices. We use them to watch TV, stream video from the Internet and to make video calls and video conferences. The video is compressed before being stored and transmitted, which makes it sensitive to errors. The errors frequently appear in wireless networks, especially in areas with poor signal coverage or in places with too many devices. Lost or corrupted packets cause subsequent macro blocks to be marked as corrupted by the decoder. The impact of transmission errors can be minimized in the receiver through error concealment.
Error concealment is an error control technique capable of mitigating the error effects on multimedia using all the available decoded data (both correctly received and erroneous). Recently published over-segmentation algorithms operating in real-time have opened new ways of error concealment for streamed videos. We introduce a fast error concealment technique where the corrupted regions are restored by texture extrapolation from the surrounding regions logically associated through image segmentation. Our method can faithfully complete the texture and edges in the missing areas. However, areas in the image containing unsharp edges or gradients, which are difficult to segment properly, are the main problem producing artifacts in the result. Therefore, in addition we propose an extended form of segmentation which adds soft edges to the segments and allows them to overlap. As result the unsharp edges and gradients are maintained in the concealed parts of the image.
The Final segmentation of an over-segmented image is usually obtained by merging neighboring regions based only on their color similarity. We propose a novel method that in addition classifies the image regions based on their texture features and we show that this improvement leads to better results.
The full thesis is available here.