Image Restoration Based on a Subjective Criterion

The problem of removing random noise from gray tone images without significantly sacrificing the subjective resolution is considered. Based on a subjective visibility function, which gives the relationship between the visibility of a unit noise and a measure of local spatial detail (spatial masking), two procedures are developed to adapt continuously the finite impulse response of a two-dimensional, noncausal, linear digital filter. At sharp transitions in the image intensity, the filter operator is strongly peaked to preserve the resolution, whereas in flat areas it is flat to effectively average out the random noise. The first procedure (S-filter) is computationally more efficient, but does not perform as well as the second method (SD-filter) which requires solution of a new optimization problem at every picture element. Results of several simulations are presented to demonstrate the feasibility of our approach. Extensions are pointed out to incorporate different adaptation procedures and psychovisual criteria other than the type of spatial masking used here.