A novel DCT domain image coder is proposed according to both intra-block clustering of significant coefficients and inter-block similarity. The algorithm utilizes morphological dilation to extract and encode the clustered significant coefficients, and elaborately adaptive arithmetic models are designed for significance encoding and signencoding, which can eliminate intra-block and inter-block correlation of the DCT coefficients effectively. The difference reduction is used to encode the position of the start pixel in each cluster and scattered coefficients, which can improve the coding efficiency of the insignificant coefficients. Furthermore, the algorithm employs pre-filtering and post-filtering to optimize coding performance and restrain blocking artifacts. As a kind of embedded coder, the algorithm is also rate scalable. Experimental results show that the performance of the new coder is superior to conventional coders. For example, for the Lena and Barbara image at 0.25 bpp, the proposed method outperforms JPEG2000 by 0.4 dB and 1.7dB in peak signal-to-noise ratio, respectively.