in this paper, a new CAVLC decoding architecture with a soft-input design concept is proposed. We introduce the soft-decision information to localize the erroneous position at macroblock (MB) levels. Specifically, we compare and select the minimal square difference between the received soft streams and decoded codewords. The corrupted MBs can be early detected and thereby concealed from neighboring pixels. Therefore, more than 1dB of PSNR gain can be achieved under the 2.7