•  
  •  
 

Corresponding Author

Yagneshkumar Jayantilal Parmar

Subject Area

Computer Science

Article Type

Original Study

Abstract

Depth images from low-cost sensors often suffer from blurred edges and structural distortions when processed with standard super-resolution models. While FSRCNN is efficient for RGB images, it struggles to handle the unique geometric requirements of depth maps. To solve this, we propose the Edge Guided Channel Attention FSRCNN (EGCA FSRCNN). This method incorporates an edge-guided modulation mechanism to preserve object boundaries and a Squeeze and Excitation (SE) block to focus on critical structural features. A major benefit of this framework is the use of frozen, pretrained FSRCNN weights, which bypasses the requirement for retraining. Our evaluation on the UTKinect, Middlebury, and NYU Depth V2 datasets confirms the effectiveness of this approach. Compared to the baseline FSRCNN, our model achieves a superior Edge Preservation Index (EPI = 0.8935), ensuring sharper structural transitions.our model provides a highly practical balance between performance and speed. While large-scale transformer models can achieve high accuracy, they require significant computational power and long inference times. In contrast, our EGCA FSRCNN maintains a small memory footprint of only 10.1 MB and delivers sharp, consistent results in real-time. These findings demonstrate that our method is a computationally efficient, lightweight solution for real-world depth enhancement tasks where edge accuracy and low computational cost are essential.

Keywords

Channel Attention, Depth Image Super Resolution, FSRCNN

Creative Commons License

Creative Commons Attribution 4.0 License
This work is licensed under a Creative Commons Attribution 4.0 License.

Download

Share

COinS