Densely Connected Feature Pyramid Network for Image Segmentation

Abstract

Image segmentation is a specific image processing technique used to segment a picture into two or more semantic regions. This paper proposes a densely connected feature pyramid segmentation network and applies it to the segmentation of images in real driving scenarios. The feature pyramid network is a kind of feature extractor, which is originally used for object detection. This paper applies it to image segmentation tasks. The densely connected network is used as a part of feature pyramid network to extract features from bottom to top. Through the lateral connection, the features of the bottom-up part and the top-down part are merged. In the final merge stage, the extracted features are transformed to the same size through upsampling and then concatenated together, and finally the segmentation map is output. The experiment is conducted on the CamVid dataset, which is a dataset in actual driving scenarios. In the experiment, the generalization ability of the segmentation model is improved through data enhancement. Based on the densely connected feature pyramid segmentation network, the F1-score on the test set is 0.8895, and the IoU-score is 0.8209.