Deep-recursive residual network for image semantic segmentation

A good semantic segmentation method for visual scene understanding should consider both accuracy and efficiency. However, the existing networks tend to concentrate only on segmentation results but not on simplifying the network. As a result, a heavy network will be made and it is difficult to deploy such heavy network on some hardware with limited memory. To address this problem, we in this paper develop a novel architecture by involving the recursive block to reduce parameters and improve prediction, as recursive block can improve performance without introducing new parameters for additional convolutions. In detail, for the purpose of mitigating the difficulty of training recursive block, we have adopted a residual unit to give the data more choices to flow through and utilize concatenation layer to combine the output maps of the recursive convolution layers with same resolution but different field-of-views. As a result, richer semantic information can be included in the feature maps, which is good to achieve satisfying pixel-wise prediction. Meriting from the above strategy, we also extend it to enhance Mask-RCNN for instance segmentation. Extensive simulations based on different benchmark datasets, such as DeepFashion, Cityscapes and PASCAL VOC 2012, show that our method can improve segmentation results as well as reduce the parameters.