Automatic enhancement of noisy image sequences through local spatiotemporal spectrum analysis

Abstract

A fully automatic method is proposed to produce an enhanced image from a very noisy sequence consisting of a translating object over a background with different translation motion. The method is based on averaging registered versions of the frames in which the object has been motion compensated. Conventional techniques for displacement estimation are not adequate for these very noise sequences, and thus a new strategy has been used taking advantage of the simple model of the sequences. First, the local spatio-temporal spectrum is estimated through a bank of multidirectional/multiscale third order Gaussian derivative filters, yielding a representation of the sequence that facilitates further processing and analysis tasks. Then, energy-related measurements describing the local texture and motion are easily extracted from this representation. These descriptors are used to segment the sequence based on a local joint measure of motion and texture. Once the object of interest has been segmented, its velocity is estimated applying the gradient constraint to the output of a directional band-pass filter for all pixels belonging to the object. Velocity estimates are then used to compensate the motion prior to the average. The results obtained with real sequences of moving ships taken under very noisy conditions are highly satisfactory, demonstrating the robustness and usefulness of the proposed method.