Target images like depth maps often suffer from blurred boundaries or noise when taken from poor quality sensors or upsampled from low-resolution images. Classical methods utilized a target sub-network and a guidance sub-network to extract features independently from the pair-wise target image and guidance image, and then a reconstruction sub-network processed the concatenated features to produce the final result. These methods only took the high-level features from the bottom layers, but neglected the inherited multi-scale attributes of the low-level vision tasks. Research has shown that features from shallow layers can also contribute to the reconstruction accuracy in single image super resolution tasks. Besides, a large receptive field can provide more structure information, which is critical for super-resolution problems. However, most existing methods only took a relative small receptive field. To deal with these problems, this paper proposed a Dense Deep Joint Network (DDJF) based on cascaded small filters and dense skip connections. First, the cascaded small filters were used to enlarge the receptive field and reduce the parameter number. Second, to alleviate the vanishing-gradient problem and strengthen the information flowing through the deep network, the dense skip connections were applied to ensure that the features of both shallow layers and deep layers can directly pass to the reconstruction layer. Finally, a residual learning strategy was applied to improve the training procedure. Experiment results by quantitative and visual comparisons with some state-of-the-art methods show the superiority of the proposed method.
