Scene-based Non-uniformity Correction using Complementary Fixed Pattern Noise Models

We propose a novel scene-based non-uniformity correction (NUC) scheme for infrared focal plane array (FPA) detectors to account for both high-frequency and low-frequency fixed pattern noise (FPN). High-frequency FPN can be significantly reduced by the recent scene based NUC algorithms. However, low-frequency FPN caused by stray light, optical effects, heat dissipation and so forth, is commonly compensated by calibration based NUC methods. In this work, we aim to reduce both the low-frequency and high frequency components of FPN by using an efficient combination of registration based and constant statistics based approaches. We exploit scene variations through the video sequence and find the underlying low-frequency noise by smartly averaging frames based on their motion and detail content. Thus, we add the generalization power of constant statistics approach to existing scene-based NUC methods to obtain lower FPN in both high-frequency and low-frequency components. The performance of the proposed method is experimented on a public dataset corrupted by real FPN and evaluated by PSNR metric in comparison to a state-of-the-art scene-based NUC method.

1. Introduction

Infrared imagery has been of interest in biomedical, military, security, autonomous driving and various other fields. The focal plane arrays (FPA) used in infrared images have a well-known non-uniformity problem in which the detector pixels have non-uniform response to a given uniform radiance. This problem is alleviated by non-uniformity correction algorithms which are evolved into two main categories: calibration-based methods [7], [13], [2], [6] and scene-based methods [11], [10], [5], [3], [9], [12], [15], [14], [1]. In general, calibration-based methods are simple and easy to implement, however, they demand the capture of known uniform surfaces which leads to interruption of the video flow. This is not very desirable especially in critical applications that do not tolerate any loss of sight of target objects. Another drawback of these methods is that they often require special and expensive tools to generate reference surfaces to accomplish the calibration based NUC process. Such limitations gave rise to scene-based approaches in which the video sequence is not interrupted and only the scene image frames are used to estimate the FPN. Possible problems with these methods are the complexity of the algorithms and ghosting effects due to erroneous FPN estimations. However, recent methods [14], [1] are capable of producing both satisfactory FPN correction and negligible or even unnoticeable ghosting artifacts.