Abstract:

In view of the problems of huge parameter calculation cost and redundant parameters in semantic segmentation tasks, this study proposes a channel feature pyramid module to solve this problem. Based on the channel feature pyramid module and a lightweight attention mechanism, a real-time semantic segmentation network is constructed. The channel feature pyramid module creates sufficient receptive field and densely utilizes context information, and gradually combines feature maps with summation operations starting from the second channel, and concatenates them to build the final hierarchical feature map, which is used in regular convolutional layers. The attention mechanism of the convolution module is added later to improve the segmentation accuracy. Without any pre-training and post-processing, the algorithm achieves a segmentation accuracy of 68.1% on the CamVid data set using only 0.75 MB parameters and 5.3 MB memory on a single GTX2080Ti, and 56 frames on the Cityscapes data set. The inference speed achieved an average interaction ratio of 75.7%.

Key words: prediction task, semantic segmentation, inference speed, channel features, attention mechanism, receptive field, context information, mean intersection over union