Author(s)

Xuejing Ding

Corresponding Author

Xuejing Ding

Abstract

In this paper, the current mainstream image semantics segmentation methods based on full convolution network are analyzed. Through analysis, it is found that FCN weakens the function of classifiable features while acquiring semantic information of the target, resulting in insufficient expression ability of the feature, resulting in insufficient accuracy of target recognition, incomplete segmentation and even loss of the target, etc. To solve this problem, an image semantics segmentation method based on multi-scale feature extraction is proposed. By constructing a full convolution network and using different scale images as input of the network, the features of different scale images are extracted. Finally, the segmented image is obtained by feature fusion. Experiments show that better image semantics segmentation can be achieved by extracting and fusing multi-scale features.

Keywords

Image semantic segmentation, Multi-scale features, Full convolution network