Pixel Voting Decoder: A novel decoder that regresses pixel relationships for segmentation

作者：

Highlights：

• The introduction of the instance segmentation and semantic segmentation tasks.

• A novel encoder-decoder to regress the pixel relationships for better segmentation.

• The proposed dynamic deconvolution uses the pixel relationships as weights.

• The matrix implementation boosts the calculation speed and reduced the memory cost.

• The proposed method outperforms the well-known methods on COCO and Cityscapes.

摘要

•The introduction of the instance segmentation and semantic segmentation tasks.•A novel encoder-decoder to regress the pixel relationships for better segmentation.•The proposed dynamic deconvolution uses the pixel relationships as weights.•The matrix implementation boosts the calculation speed and reduced the memory cost.•The proposed method outperforms the well-known methods on COCO and Cityscapes.

论文关键词：Convolutional neural network,Dynamic deconvolution,Encoder-Decoder,Image segmentation,Pixel voting,Residual block

论文评审过程：Received 1 May 2021, Revised 13 September 2021, Accepted 19 December 2021, Available online 31 December 2021, Version of Record 13 January 2022.

论文官网地址：https://doi.org/10.1016/j.eswa.2021.116438