Efficient Sampling of Bayer Pattern for Long Range Small Target Detection in Color Images

Detecting small moving targets at long ranges in visible spectrum images is a challenging problem. The task typically involves using input imagery captured with a low-cost single-sensor camera equipped with a color filter array, such as a Bayer pattern. The main detection challenges include low signal-to-noise ratio, low target contrast, sensor noise, and background clutter. Biologically inspired vision based computational models have previously shown robust and adaptive performance in simultaneously overcoming these challenges. This paper presents several sampling strategies of the Bayer pattern images for the efficient construction of input images to the small target detector. An optimal sampling strategy is determined through experimental analysis of the detection performance versus the efficiency of the model. Extensive experiments are conducted using real-world, high frame rate, and high bit-depth visible spectrum image sequences containing real physically small targets, such as aerial drones, at long ranges (>2 km). Our results show that with efficient sampling, improved detection performance is achieved by using only half of the computations.