A　novel　method　of　texture-image　classification　with　rotation-ill　variance　in　DCT　compressed-domain　is　proposed.　By　reordered　the　DCT　coefficients　for　multi-resolution,　the　textural　features　are　extracted　in　DCT　transformed-domain　and　these　textural　features　can　be　used　to　classify　texture-images　in　RLE　bit-stream　directly　with　rotation-invariance.　The　experimental　results　demonstrate　that　texture-images　can　be　classified　at　higher　correct　classification　rate.