Image Semantic Segmentation Using Deep Convolutional Nets, Fully Connected Conditional Random Fields, and Dilated Convolution

摘要

Deep convolutional neural networks (DCNNs) have recently demonstrated state-of-the-art performance in advanced vision tasks, such as image classification and object detection. This work focuses on solving image semantic segmentation tasks. First, we combine a new feature extraction network with a dilated convolution layer to improve the accuracy of the model's mission. Second, we introduce multi-scale feature fusion technology to improve the performance of DCNN. Third, we combine the DCNN with fully connected conditional random field to overcome the inaccurate positioning of DCNN and optimize their output. Our approach is demonstrated on the PASCAL VOC-2012 Image Semantic Segmentation dataset, where 78.1% IOU accuracy is achieved in the test set. Our approach can compute neural network responses intensively at 9 frames per second on modern GPUs.