Strong-Structural Convolution Neural Network for Semantic Segmentation

Abstract

We present a combinatorial deep convolutional neural network architecture, termed strong convolution neural network (SSN), for semantic segmentation task. The structure of SSN consists of two components: Increment feature convolution neural network and post-process Conditional Random Fields unit (CRFs). The increment feature CNN unit has three parts: I-Block, Deconvolution layer and Transition Block. I-Block employs increment convolution to efficiently maintain feature information. Before passing through pooling layer, we put the feature map into activate layer ReLU, and batch normalization layer. In Decoding stage, we use skip-connects to keep the pooling index information. To enforce the correlation of same semantic labels, we define the strong semantic label (SSL) stage to intensify the pairwise potential energy. To achieve high computation performance, we make further improvement on SSL by employing the adaptive soft semantic sections label method. We proposed the adaptive strong semantic label selection algorithm to generate the SSL. Through the CRFs unit, with unitary energy and pairwise edge energy, the semantic segmentation initial labels transform semantic segmentation labels. Experimental evaluation reveals the training time versus accuracy trade-off involved in achieving good segmentation performance.