基于DeepLabV3+与注意力机制相结合的图像语义分割

Image Semantic Segmentation Based on Combination of DeepLabV3+ and Attention Mechanism

基于DeepLabV3+进行图像分割时,在特征提取阶段忽略了不同级别的特征图中存在的特征重要程度不同,丢失了大量的细节信息,致使分割效果不佳。针对该问题,提出了一种基于DeepLabV3+与注意力机制相结合的图像语义分割算法。在骨干网络Xception模型中提取两条低级特征作为解码器的输入特征,提高特征提取的准确性;采用通道注意力模块有效融合高级特征,获取丰富的上下文信息;采用空间注意力模块提取低级特征,过滤背景信息,减少细节信息的丢失;采用深度可分离卷积代替空洞卷积有效降低参数量,提高计算速度;同时,采用焦点损失作为损失函数通过降低内部加权,提高最终的分割效果。实验结果表明,所提算法在PASCAL VOC 2012数据集上的平均交并比(mIoU)值达到了84.44%,与传统算法和基于DeepLabV3+改进的算法相比,有效提高了特征提取的准确性,减少了特征细节信息的损失,对最终的分割效果有了较好的提升。

In image segmentation based on DeepLabV3+, the different importance of features in different levels of feature images are ignored in the feature extraction stage, and a large amount of details are lost, resulting in poor segmentation effect. To solve this problem, an image semantic segmentation algorithm based on the combination of DeepLabV3+ and attention mechanism is proposed. Two low-level features are extracted in the backbone network Xception model as input features of the decoder to improve the accuracy of feature extraction.The channel attention module is used to effectively integrate high-level features and obtain rich context information. The spatial attention module is used to extract low-level features and filter background information to reduce the loss of details. The depthwise separable convolution is substituted for void convolution to effectively reduce the amount of parameters and improve the calculation speed. At the same time, the focus loss is used as the loss function to improve the final segmentation effect by reducing the internal weighting. Experimental results show that the mean intersection over union(mIoU) value of the proposed algorithm on PASCAL VOC 2012 dataset reaches 84.44%. Compared with the traditional algorithm and the improved algorithm based on DeepLabV3+, the proposed algorithm effectively improves the accuracy of feature extraction, reduces the loss of feature details, and improves the final segmentation effect.