Backbone Feature Enhancement and Decoder Improvement in HRNet for Semantic Segmentation

Addressing issues such as the tendency for small-scale objects to be lost, incomplete segmentation of large-scale objects, and overall low segmentation accuracy in existing semantic segmentation models, an improved HRNet network model is proposed. Firstly, by introducing multi-branch deep stripe convolutions, features of multi-scale objects are adaptively extracted using convolutional kernels of different sizes, which not only enhances the model’s ability to capture multi-scale objects but also strengthens its perception of the contextual environment. Secondly, to optimize the feature aggregation effect, the axial attention mechanism is adopted to aggregate image features along the x-axis and y-axis directions respectively, effectively capturing long-range dependencies within the global scope, and thus achieving precise positioning of objects of interest in the feature map.Finally, by implementing the progressive fusion-based upsampling strategy, it facilitates the complementary fusion of semantic information and detailed information between adjacent feature maps, thereby enhancing the model’s capability to restore fine-grained details in images. Experimental results demonstrate that on the PASCAL VOC2012+SBD dataset, the mean Intersection over Union (mIoU) of the improved HRNet S model in segmenting lower-resolution images is increased by 1.54% compared to the baseline method. Meanwhile, the improved HRNet L model achieved a 3.05% increase in mIoU compared to the original model when handling higher-resolution image segmentation tasks on the Cityscapes dataset, and attained the highest segmentation accuracy in 15 out of the 19 different scale classification categories on this dataset.These results indicate that the proposed method not only exhibits high segmentation accuracy but also possesses strong adaptability to multi-scale objects.

@article{Feng2024,
  title     = {Backbone Feature Enhancement and Decoder Improvement in HRNet for Semantic Segmentation},
  journal   = {International Journal of Advanced Computer Science and Applications},
  volume    = {15},
  number    = {10},
  year      = {2024},
  publisher = {The Science and Information Organization},
  author    = {HanLei Feng and TieGang Zhong},
  doi       = {10.14569/IJACSA.2024.0151098},
  url       = {https://doi.org/10.14569/IJACSA.2024.0151098}
}