At present, most JPEG steganographic methods embeded secret message by changing the value of DCT coefficients. The embedding method was bound to affect the dependencies existed in DCT coefficients. So it was necessary to make use of the dependencies existed in DCT coefficients in steganalysis. Co-occurrence matrices were a common tool to extract the dependencies existed in DCT coefficients. The coefficients must be scanned before extracting features using this tool. The most JPEG steganalysis methods used line scanning or column scanning and studied less on other scanning patterns. So they didn’t make the most of the dependencies existed in DCT coefficients. In view of this problem, it presented an novel JPEG steganalysis based on coefficient dependencies in the paper. It analyzed the inter-block and intra-block depenmencies along different scanning patterns existed in DCT coefficients. High orders co-occurrence matrices were applied to extract different orientation and distance dependencies existed in inter-block and intra-block DCT coefficients as features. The experiments on four steganographic methods which had eleven embedding rate showed the detection accuracy of this method was improved obviously compared with similar low dimensional feature method, and this method could effectively control the feature dimension without significantly reducing the detection accuracy compared with similar high dimensional feature method.