Fast colorization for single-band thermal video sequences

Abstract

The color night-vision technology is advantageous to scene interpretation by displaying original monochrome night-vision imagery with colors, leading the future trend of night-vision development. Until recently, the methods of color night-vision are mostly based on multi-band sensor fusion that do not apply to thermal systems which are only sensitive to one spectral band. To address this, we present a novel technique in this paper aiming to directly give thermal videos day-time color appearance. We observe that the interframe pixel׳s luminance changing over thermal videos approximately follows a Double Lomax distribution. Based on this observation and computational reasons, we propose a star-shape mask for pixel matching. We further show how the mask can be used for efficient thermal video colorization. Additionally, considering the accumulating errors in sequence, we propose a simple region-similarity based abnormal color correction mechanism. Experiments on real applied thermal night-vision videos show that our method can efficiently render single-band thermal sequence with correct colors at 12–33 frame-per-second (fps), providing observers with enhanced facility of surrounding recognition and target recognition.