Aircraft Detection from Remote Sensing Images using Single Shot Scale-invariant Face Detector with Viridis Saliency Map

Aircraft detection from remote sensing images is a useful task and essential in many ways, such as surveillance, tracking, and information collecting - however, it is still a challenging task because of a complicated of background, and a small scale of the aircraft compared to the remote sensing image size. Many researches aim to improve detection accuracy, but due to the problem mentioned above, it makes many proposed methods perform less effectively. Therefore in this paper, we proposes a practical saliency map that can significantly reduce noise from the background by using Viridis colormap, Gaussian blur, Otsu's thresholding method, and closing morphological operation to reduce noise from the background and create a binary image. Then, we find a contour of all objects in the binary image and fill a hole inside them to get the better binary image. After that, we combine a grayscale image with the binary mask to create the saliency map. Finally, we use the four channels image, which composes of three channels from an RGB input image and one channel from the saliency map, as the input for training with the CNN model based on the (SFD)-F-3 network. The experimental results show that applying the saliency map and the RGB input image can improve the recall rate from 94.99% to 98% compared to using only the RGB input image with the (SFD)-F-3 network. Thus this method performs well and sufficient to detect the aircraft in the high complexity of the background from remote sensing images.