基于区块自适应特征融合的图像实时语义分割

近年来结合深度学习的图像语义分割方法日益发展, 并在机器人、自动驾驶等领域中得到应用. 本文提出一种基于区块自适应特征融合(Block adaptive feature fusion, BAFF) 的实时语义分割算法, 该算法在轻量卷积网络架构上, 对前后文特征进行分区块自适应加权融合, 有效提高了实时语义分割精度. 首先, 分析卷积网络层间分割特征的感受野对分割结果的影响, 并在跳跃连接结构(SkipNet) 上提出一种特征分区块加权融合机制; 然后, 采用三维卷积进行层间特征整合, 建立基于深度可分离的特征权重计算网络. 最终, 在自适应加权作用下实现区块特征融合. 实验结果表明, 本文算法能够在图像分割的快速性和准确性之间做到很好的平衡, 在复杂场景分割上具有较好的鲁棒性.

Real-time Image Semantic Segmentation Based on Block Adaptive Feature Fusion

Recently, image semantic segmentation has made great progress with deep learning, which benefits robotics and automatic driving vehicle. This paper proposes a real-time semantic segmentation algorithm based on block adaptive feature fusion (BAFF). Under the framework of a light convolutional network, a block adaptive feature fusion algorithm is proposed in the context-embedding module, to improve the accuracy of real-time semantic segmentation. First, the problem caused by the different size of receptive field in layers is analyzed, and a feature fusion mechanism with block weight is presented on SkipNet. Then, layers' feature integration is carried on by three-dimension convolution. The feature-weights are calculated by an additional network with depthwise-separable-convolutions (DSC). Finally, the features are fused under adaptive weights. Experiments show that this method obtains excellent segmentation results with a good balance between rapidity and accuracy and owns robustness on segmentation of complex scenes.