HiFused-Depth: High-Frequency Fused Depth Estimation with Knowledge Guidance

Monocular depth estimation (MDE) is a fundamental task in computer vision with a wide range of applications. However, achieving accurate depth prediction remains challenging in scenes with complex structures and low texture areas. To address this issue, we propose a new depth estimation network that integrates high-frequency information enhancement and knowledge extraction. Specifically, the Sobel operator is used to extract high-frequency features such as edges and textures from the input RGB image, and merge them into the feature representation to enrich geometric understanding. Establish a high-frequency perceptual loss function to guide the network to focus on texture defects and structurally blurry areas, improving depth prediction accuracy. In addition, adopting a knowledge distillation strategy from a pretrained external teacher network, which transfers both global semantic context and fine-grained geometric cues to the student model, enhances generalization ability in different environments. Extensive experiments conducted on the NYUv2 and KITTI datasets have shown that the proposed method consistently outperforms existing methods in multiple standard evaluation metrics, achieving up to 7.3% reduction in log 10 error and 6.3% reduction in Abs Rel on NYUv2, and 6.9% reduction in Sq Rel on KITTI compared with the strong NeWCRFs baseline.