多尺度特征自适应融合的图像语义分割算法

Image Semantic Segmentation Algorithm Based on Adaptive Fusion of Multi-scale Features

为解决Deeplab v3+网络在解码过程中仅融合一个尺度编码特征,导致部分细节信息丢失,从而造成最终分割结果较为粗糙等问题,提出多尺度特征自适应融合的图像语义分割算法.该算法在Deeplab v3+的解码过程中使用自适应空间特征融合结构,给不同尺度的编码特征分配自适应的融合权重,通过融合编码过程中的多尺度特征进行特征图的上采样,从而实现图像更加精细的语义分割效果.实验结果表明,该算法在Cityscapes数据集上达到了95.05%的像素准确度和69.36%的平均交并率,且对大部分小尺度目标物体的分割更为精准;在Vaihingen遥感图像数据集上本文提出的算法达到了83.49%的像素准确度和68.77%的平均交并率,进一步验证了本文改进算法的泛化性.

In order to solve the problem that the Deeplab v3+only fuses one scale encoding feature in the decoding process,which leads to the loss of some detailed information and causes the final segmentation result to be rougher,a image semantic segmentation algorithm based on adaptive fusion of multi-scale features is proposed.The algorithm uses an adaptive spatial feature fusion structure to assign adaptive fusion weights to encoding features of different scales in the decoding process of Deeplab v3+,and upsamples the feature map by fusing the multi-scale features in the encoding process to achieve a more refined image semantic segmentation result.Experimental results show that the algorithm achieves 95.05%pixel accuracy and 69.36%mean intersection over union on Cityscapes,and segmentation of most small-scale target objects is more accurate.On the Vaihingen dataset the proposed algorithm achieves a pixel accuracy of 83.49%and a mean intersection over union of 68.77%,which further verifies the generalization of the algorithm.