Real Time Semantic Segmentation Algorithm Via Attention Based Feature Fusion

Deep learning based computer vision algorithms are widely developed in recent years due to the availability of a large number of image data sets.Beyond image-level classification and recognition,state-of-the-art methods can also perform pixel-level localization and semantic labeling,facilitating a large variety of applications such as intelligent driving,medical diagnosis,face segmentation and recognition and other fields.Driven from both academic research and industrial potential,deep learning based semantic segmentation has attracted ever-increasing research attention in the past decade.Most existing semantic segmentation algorithms improve the model accuracy from the perspective of expanding receptive field,such as using spatial pyramid module to expand receptive field.However using hole convolution with large void ratio will slow down the calculation speed,while the smaller void ratio will lead to too small receptive field.To solve this problem,we propose a smooth version of spatial pyramid module,which expands the receptive field without increasing the amount of computation,and ensures real-time model performance.At present,the cutting-edge semantic segmentation networks directly performs Concat or Add operations on two feature maps when doing feature fusion.Since the two feature maps with different resolutions are quite different,directly weighting the two feature maps will cause serious accuracy drop.To handle this problem,in this thesis we propose a feature fusion module based on the attention mechanism to fully integrate shallow spatial information and deep semantic information,and improve the accuracy of the semantic segmentation model.A series of experiments were carried out on Cityscapes,VOC2012 and Sun-RGBD data sets.The experimental platform is NVIDIA-1080Ti graphics card,in which 73.75%MIOU is achieved on Cityscapes data sets,and the speed is 89.29 FPS.On the Sun-RGBD data sets,58.81%MIOU was obtained for thirteen classifications,71.42%MIOU was obtained for five classifications,and the speed was 76.92 FPS.The experimental results show that the proposed real-time semantic segmentation algorithm using attention-based mechanism feature fusion achieves better accuracy and real-time inference speed on benchmark data sets.