A small object detection method based on CNN in serial multispectral images

This paper introduces a small object detection method based on 3D CNN in multispectral serial images. The temporal dimension is set as 5, which means the convolutional filter is n*n*c*5, because 5 consecutive images are sufficient to remove noises. We proposed a method for corresponding training sample preparation, where the true small object region is calculated and object moving in every direction is simulated which is called object simulation base region. The negative sample set is composed of random multispectral images while the positive sample set is composed by linear superposition of object simulation base regions. Experimental results proved this method is feasible and effective.