一种粗细特征增强语义分割算法

A Coarse-fine Feature Enhanced Semantic Segmentation Algorithm

针对现有语义分割网络由于特征提取能力不足导致漏分割和过分割的问题,提出一种粗细特征增强语义分割算法。该算法主要由粗特征提取模块、细特征提取模块和特征融合模块组成,两个特征提取模块分别获取表示颜色、形状的粗特征和表示边缘、角点的细特征,在特征融合模块中将二者结合,可以获得上下文联系更加密切、表达更全面的特征信息,进而提升语义分割精度。在Pascal VOC2012数据集上的实验结果表明,粗细特征增强语义分割算法优于现有同类算法,与代表性的DeepLabv3+网络相比较,平均精度(mIoU)提高了0.66%。

For the problem of missing and over-segmentation due to insufficient feature extraction capabilities in semantic segmentation networks,a coarse-fine feature enhanced semantic segmentation algorithm is proposed.The algorithm is mainly composed of a coarse feature extraction module,a fine feature extraction module and a feature fusion module.The two feature extraction modules obtain coarse features representing colors and shapes and fine features representing edges and corners,respectively.They are combined in the feature fusion module to obtain more contextual and comprehensive feature information to improve the semantic segmentation accuracy.Experimental results on the Pascal VOC2012 dataset show that the proposed method is better than the existing similar algorithms.Compared with the representative DeepLabv3+network,the average accuracy(mIoU)is improved by 0.66%.