Super-resolution segmentation of blurry pointer meter images based on frequency supervision

Accurate segmentation of pointer meter images is crucial for efficient meter reading tasks. However, the accuracy of meter segmentation is often hindered by image blurriness due to resolution discrepancies, an issue that existing methods have largely overlooked. This paper addresses this gap by introducing a novel super-resolution semantic segmentation network (SegSRCNN) designed to reconstruct blurry images and enhance segmentation accuracy by leveraging the synergies between super-resolution (SR) and segmentation tasks. The SegSRCNN architecture features a multi-scale feature encoder and an SR fusion decoder that dynamically establish feature associations between the two tasks. We also incorporate a Channel-attention and Self-attention Mixed Block (CSMB) within both the encoder and decoder, integrating channel and self-attention mechanisms to optimize local feature relationships and global contextual awareness, resulting in improved segmentation outcomes. Additionally, we propose a novel training objective, frequency supervision, which emphasizes high-frequency consistency, significantly boosting the model’s ability to recover fine details. Experimental results show that SegSRCNN achieves a mIoU of 85.43% on the pointer meter dataset, surpassing SOTA methods by at least 0.32%. These findings highlight the unique contributions of our approach in the realm of pointer meter image processing. • Propose a super-resolution segmentation network, SegSRCNN, to enhance pointer meter images segmentation. • Propose CSMB module, which can optimize both local feature relationships and global contextual awareness. • Propose a novel a frequency-based training objective, which can improve detail recovery in blurry images. • SegSRCNN enhances the segmentation accuracy and reading performance for low-resolution blurry meter images.