Steerable Pyramid Denoising as Preprocessing for 2D-Object Detection and Visualization of Tomographic Reconstructions

Abstract

In electron microscope images of biological macromolecules the information on their structure is not directly accessible since most of the signal is buried in noise. A common approach for signal reconstruction is to average over a large number of identical structures. A crucial point is the detection of this structures, requiring a detection algorithm being able to perform well in images with a low signal-to-noise ratio (0 dB), a low and varying contrast, in-plane rotated and densely distributed objects in the scene, strong object deformations, and the possible occurrence of false class objects and object agglomerations. We propose a rotation invariant and robust multiple 2D-object detector which consists of (a) a steerable pyramid based denoising and (b) a constrained circular harmonic function filter. Experiments on our imagery show that denoising prior to the actual detection significantly reduces the number of erroneous findings, thus increasing the reliability of object detectors. When imaging unique, non-repeatable structures signal reconstruction based on averaging is not possible. We propose the application of steerable pyramid denoising to electron tomographic reconstructions. We elucidate a method to quantify the loss of information due to denoising artifacts on data with an unknown signal-noise relationship, and propose a choice for denoising parameters. Our experiments prove steerable pyramid denoising to perform well in reconstructing signal out of noisy data while preserving most of the actual information.