用于场景分割的改进DeepLabV3+算法

Improved DeepLabV3+Algorithm for Scene Segmentation

为了提升室外场景下语义分割的精度,提出一种改进的DeepLabV3+神经网络分割算法。其主干部分采用分组的ResNest网络,使各类目标训练权重占比不同,以密集连接的方式改进空洞空间卷积金字塔池化(ASPP)模块,在不牺牲特征空间分辨率的同时扩大感受野,并且提升特征复用效率。解码端融合编码端提取的3种不同尺度的低层语义特征,以恢复在降采样过程中丢失的空间信息。实验结果表明,在CityScape数据集的检测中,该算法不仅提高了目标的分割准确率,而且对全场景理解和细节处理能力均有明显提升。

In order to improve the accuracy of semantic segmentation in outdoor scenes an improved DeepLabV3+neural network segmentation algorithm is proposed.The backbone part adopts a grouped ResNest network thus the training weights of various types of targets account for different percentages.Then the Atrous Spatial Pyramid Pooling(ASPP)module is improved by using densely connected mode to expand the perceptual field without sacrificing feature spatial resolution and to improve feature reuse efficiency.The decoding side fuses the low-level semantic features extracted from the encoding side at three different scales to recover the spatial information lost in the downsampling process.The experimental results show that the algorithm not only improves the segmentation accuracy of the target in the detection of CityScape dataset but also has a significant improvement in both full-scene understanding and detail processing ability.