SACA: Sub-Sampling Aware Channel Attention for Efficient Semantic Segmentation Model Compression

In this paper, we introduce a novel sub-sampling aware channel attention (SACA) based filter pruning technique for reducing the complexity of semantic segmentation models including the U-Net and DeepLabV3+. The proposed SACA based method is considered the first technique to incorporate a channel attention with a sub-sampling awareness capability that aims at improving the recognition of the significance of convolutional layer filters in the models by attending to the most significant sub-sampling operation. The proposed optimization method introduced pruned semantic segmentation models with enhanced dice score coefficient and recall as compared to benchmark magnitude-based (e.g., <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$l1$</tex>-norm) and channel attention-based filter pruning techniques.