Computer Science · Artificial Intelligence
ESNet: An Efficient Real-time Semantic Segmentation Network
Renping Xie, Cong He, Ming Tao, Kai Ding
See on the mapAbstract
Efficient image segmentation algorithms are critical in computer vision, as they maintain high processing speeds while handling large amounts of data and providing practical solutions in resource-limited environments. While existing classical segmentation methods achieve good results, their real-time performance can be further improved. To address this issue, we propose an efficient segmentation network (ESNet) to capture extensive contextual information and improve segmentation fineness, which are two main issues in semantic segmentation, while ensuring real-time capability. Then, we first introduce an efficient context module (ECM) that effectively enlarges the effective receptive field (ERF) of the model and improves its performance in integrating contextual information. Subsequently, a skip connection with a simple feature fusion module (SFFM) is designed to provide rich detail information, thereby improving segmentation fineness. The efficacy of ESNet was evaluated on the PASCAL VOC2012 semantic segmentation dataset against several state-of-the-art methods. ESNet achieves 71.4 FPS at a resolution of 1024×2048 on a single NVIDIA GeForce RTX 3090 GPU and achieves 62.42% mIoU at a resolution of 320×480 with an extremely simple training recipe, striking a fine balance between segmentation accuracy and inference speed.