We propose a new approach to adaptive image regularization based on a neural network, hierarchical cluster model (HCM). The HCM bears a close resemblance to image formation. Its sparse synaptic connections are effective in reducing the computational cost of restoration. We attempt to achieve adaptive restoration by assigning entries of a novel, optimized regularization vector to each homogeneous image region. The degraded image is first segmented and partitioned into smooth, texture, and edge clusters. It is then mapped onto a three-level HCM structure. An evolutionary scheme is proposed to optimize the regularization vector by minimizing the HCM energy function. The scheme progressively selects the well-evolved individuals that consist of partially restored images, their corresponding cluster structures, segmentation maps, and the optimized regularization vector. Experimental results show that the new approach is superior in suppressing noise and ringing at the smooth background while effectively preserving the fine details at the texture and edge regions. An important feature of the method is that the empirical relationship between the optimized regularization vector and the local perception measure can be reused in the restoration of other degraded images. This generalization removes the overhead of evolutionary optimization, thus rendering a very fast restoration.