Semantic Image Segmentation With Propagating Deep Aggregation

In this article, we propose a new architecture of feature aggregation, which is designed to deal with the problem that the information of each convolutional layer cannot be used reasonably and the shallow layer information is lost in the process of transmission. In this method, propagating deep aggregation (PDA) is transplanted into the DeepLab-ASPP model to generate a new model combining structural continuity with feature aggregation. This new model can be divided into several stages according to the output resolution. In these stages, more possible feature fusion combinations can be realized, and the result from aggregation propagated among stages can be optimized and updated to get the best results. We demonstrate the effectiveness of the proposed model on the PASCAL VOC 2012 data set and the PASCAL-Context data set. Our method achieves state-of-the-art performance on two public benchmarks and significantly outperforms the previous results, 78.8% (versus 62.4%) on the PASCAL VOC 2012 data set and 46.2% (versus 37.8%) on the PASCAL-Context data set.

I. Introduction

Image semantic segmentation is one of the most basic techniques in computer vision. It is to assign a semantic label to each pixel in the image and divides adjacent pixels with the same semantic label into the same region. Each region is independent and has the same attributions belonging to the same object class, such as color, shape, quality, and illumination. Semantic segmentation is the most basic and critical technology in object recognition, video data monitoring, and other subsequent processing. It plays an important role in road and building measurement, urban planning, and road monitoring. However, image semantic segmentation is still a very challenging task in road detection. There are some common problems in semantic segmentation [1], [2], such as the mismatch of object relations, the misjudgment of confusing object categories, and the neglect of nonobvious categories, which are affected by global background information partially. These problems make the classification of image pixels more difficult, which causes the prediction result to deviate from the ground truth.