서강대학교

초록/요약

Recently, there have been many studies using convolutional neural network (CNN) to improve the performance of various image processing or computer vision applications. According to development of CNN, networks with high performance such as generative adversarial networks (GANs) have been presented. We propose an RGBD saliency detection method using conditional GANs (cGANs). The proposed networks have two-stream RGBD saliency detection generator that generates the pixel-wise saliency map. The final saliency map is generated by refining the pixel-wise saliency map using the guided filtering. However, in the case of RGBD saliency detection, it is difficult to improve the detection performance using CNN because of the lack of the training set. In this thesis, by generating synthetic depth maps to train the cGANs for the RGBD saliency detection, we can improve the performance of the saliency detection. Training using the generated synthetic and real depth maps is efficient for over-fitted cases and enhances the overall performance. In addition, quantitative and qualitative comparisons on two well-known datasets show that the proposed method has higher performance than existing RGBD saliency detection methods.