CCMFRNet: A Real-Time Semantic Segmentation Network with Context Cascade and Multi-scale Feature Refinement

At present, many of the popular semantic segmentation networks focus on accuracy and require a lot of computational overhead, which results in a very slow inference speed and is difficult to deploy in practical application scenarios. On the other hand, many works sacrifice the performance of segmentation networks in pursuit of real-time inference speed. Therefore, in semantic segmentation, balancing accuracy and real-time performance becomes a formidable challenge. For this challenge, we propose a lightweight semantic segmentation network that takes into account both accuracy and real-time performance, named CCMFRNet. The core components of CCMFRNet are the Context Cascade Module (CCM) and the Multi-scale Feature Refinement Module (MFRM). CCM consists of three Dense Cascade Dilated Convolution Modules (DCDM), which are cascaded in a short-term dense cascade, aiming to obtain rich multi-scale context information to enhance information representation. MFRM adopts the attention mechanism to realize deep features to guide the captured shallow multi-scale spatial features. It aims to capture high-quality and multi-scale shallow features to enrich the feature space and more effectively refine the spatial details information. The proposed method achieves an accuracy of 72.6% MIoU at speed of 32 fps on Cityscapes test datasets.