Author(s)

Wen Chen, Shuhao Ma, Jin Wang, Qing Zhu

Corresponding Author

Wen Chen

ABSTRACT

Image pyramid is an significant approach to improve the performance of the convolutional neural networks (CNNs) on the computer vision tasks(e.g., image classification, object detection, image segmentation ), but it takes huge computation time and occupies large memory. In this paper, image pyramid has been improved. We first point out that different tasks and different data sets have different requirements for feature fusion, and then we propose a more flexible approach, called Deep Feature Pyramid (DFP), which can alleviate the large memory requirements and a lot of computation time of image pyramid to some extent. Our approach enables the model to achieve good performance with less memory overhead and computation time. Compared with image pyramid, our method greatly reduces the memory occupancy and computation time, especially when the image scale is large. We validate the performance of DFP in semantic segmentation, which is a very demanding task for feature fusion, and the experimental data set is Pascal VOC2012.

KEYWORDS

Deep convolutional neural networks, semantic segmentation, image pyramid, multi-scale feature fusion, deep feature pyramid, pascal voc dataset