全局特征提取的全卷积网络图像语义分割算法

以全卷积神经网络为基础设计图像语义分割算法框架,设计全局特征提取模块提升高维语义特征的提取能力,引入带孔卷积算子保留图像细节并提升分割结果的分辨率。通过搭建端到端的图像语义分割算法框架进行训练,在可见光数据集上对算法框架进行性能评估,结果表明,本文方法在可见光图像上取得良好的语义分割性能和精度。本文还在不借助红外数据标注训练的情况下对红外图像进行分割,结果证明本文方法在典型红外目标如行人、车辆的分割中也有较好的表现。

Image Semantic Segmentation Based on Fully Convoluted Network with Global Feature Extraction

We employed a fully convoluted network to perform the image semantic segmentation task.In detail,we introduced a global feature extraction module to enhance the high-level semantic feature extraction ability.Furthermore,we adopted the dilate convolution operation to preserve image details and increase the resolution of prediction results.We evaluated and analyzed our end-to-end semantic segmentation algorithm on visible image datasets.The results demonstrated that our proposed approach achieved a satisfactory accuracy and better visual effect.We also evaluated our framework on infrared images without training of semantic labels.The results have shown that our algorithm can obtain significance visualization on classical objects segmentation such as humans and cars.