一种融合多级特征信息的图像语义分割方法

卷积神经网络因为其强大的学习能力，已经在语义分割任务中取得了显著的效果，但是如何有效地利用网络在浅层次的视觉特征和深层次的语义特征一直是研究的热点，以此为出发点，提出了一种融合多级特征信息的图像语义分割方法。通过空洞卷积提取各层级的特征，并不断迭代深层特征来丰富低级视觉信息，最后与高级语义特征合并融合，得到精细的语义分割结果。实验在PASCAL VOC 2012数据集上与主流的五种方法进行了比较，在GTX1080Ti的环境下该方法与其中性能第二的模型mIoU（mean intersection-over-union）值相比提高了2.1%，与其中性能第一的模型mIoU值仅相差0.4%，表明该方法能有效利用多层级的特征信息，实现了图像语义分割的目的。

Semantic segmentation method integrating multilevel features

Convolutional neural network has achieved remarkable results in semantic segmentation tasks, because of its powerful learning ability. However, how to effectively use the low-level visual features and high-level semantic features of the network has been a research hotspot. Therefore, this paper proposed a semantic segmentation method that integrated multilevel feature information. This method extracted the features of each level through atrous convolution, iterated the deep features to enrich the low-level visual information, and finally merged with the high-level semantic features to obtain the fine semantic segmentation results. On the PASCAL VOC 2012 dataset, this method was compared with the five main methods. In the GTX1080Ti environment, the value of mIoU(mean banding over union) increased by 2.1% compared with the model whose performance was the second, and was only 0.4% lower than that of the model whose performance was the first. The result indicates that the proposed method can effectively make use of multi-level feature information, and the realize the purpose of image semantic segmentation.