HyperSegmenter: Reappraising the potential of large kernel CNN architecture in efficient semantic segmentation

Semantic segmentation aims to precisely delineate the semantic content of each pixel in images, providing profound comprehension and precise localization for diverse vision tasks. While the advent of the Vision Transformer architecture has significantly propelled the field forward, these approaches still encounter challenges such as local inductive bias and elevated time complexity stemming from self-attention mechanisms. Addressing these issues, this paper reassesses the convolutional neural network architecture. We introduce an efficient convolutional operator and establish the SCU module as foundational to alleviate constraints observed in current methodologies. Furthermore, to mitigate redundancy within decoder structures, we endeavored to redesign a ’Sandwich’ decoder integrating the LKD and AKConv modules, specifically designed for demanding semantic segmentation tasks. Our model, termed HyperSegmenter, endeavors to enhance both efficiency and adaptability. HyperSegmenter is categorized into four iterations: Tiny, Small, Base, and Large, and underwent rigorous evaluations across three benchmark datasets-ADE20K, Cityscape, and COCO-Stuff. Experimental outcomes demonstrate substantial performance gains, achieving respective accuracies of 52.23 %, 82.54 %, and 48.91 %. These results underscore its efficacy and applicability in intricate scenarios.