The encoding speed of vector quantization (VQ) is an important problem for VQ's practical applications. Because a k-dimensional (k-D) vector can also be mathematically viewed as a k-element set so that the statistical analysis methods can be directly applied to k-D vectors. In order to approximately measure the difference between two k-D vectors, by using the well-known statistical features of the sum and the variance of a k-D vector first, the IEENNS method (S. Baek, et. al., IEEE Signal Processing Letters, vol.4, pp.325-327, 1997) has been proposed to reject most of unlikely codewords for a certain input vector. Then, by dividing a k-D vector in half to generate its two corresponding (k/2)-D subvectors and then apply the IEENNS method again to each of the subvectors, a complete-version SIEENNS method (J.S. Pan, et. al., IEEE Trans. Image Processing, vol.12, pp.265-270, 2003) has been proposed as well. Because the SIEENNS method still has a large memory and computational redundancy, a simplified-version enhanced ESIEENNS method (Z. Pan et. al., 2005 International Symposium on Circuits and Systems, pp.6332-6335, 2005) is reported recently. However, all of these subvector-based previous works just fixedly constructed its two subvectors for simplicity, which cannot guarantee a very high search performance. Instead, this paper proposes to dynamically construct the two subvectors more efficiently based on a criterion of |S yi,f-Syi,s ⇒ max by offline analyzing the property of a codeword yi. Experimental results confirmed that the proposed DESIEENNS method can improve the total search efficiency to 79.9%-88.7% compared to the latest ESIEENNS method for various input images.