Infrared image-based remote rescuee detection method in maritime situation utilizing a deep learning network and data augmentation

Abstract

In this thesis, a fast and robust infrared remote target detection network is proposed based on deep learning. Furthermore, we construct our own IR image database imitating humans in remote maritime rescue situations using the FLIR M232 IR camera. First, data augmentation is performed by applying domain adaptation from the game data to the real data. By conducting target-background separated domain adaptation, the shapes and locations of the objects maintain and only adapt the data distribution, leading to a proper algorithm for the small objects. Second, multi-scale feature extraction is performed combined with fixed weighted kernels and convolutional neural network layers. By utilizing fixed-weight kernels, background, and noise which are uniform in general are removed and enhance the gradient at the direction of the center of the objects, increasing signal-to-noise ratio. Lastly, the feature map is mapped into a likelihood map indicating the potential locations of the targets. Experimental result reveals that the proposed method can detect remote targets even under complex backgrounds which existing algorithms fail to detect. Moreover, the network trained on augmented data with game shows higher performance than that with only real data.