TY - GEN
T1 - Hybrid Feature Edge Enhancement for Self-Supervised Monocular Depth Estimation in Endoscopic Scenes
AU - Guo, Jiadong
AU - Niu, Ke
AU - Feng, Xue
AU - Li, Heng
AU - Ou, Mingyang
AU - Liu, Zeyun
N1 - Publisher Copyright:
© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2027.
PY - 2027
Y1 - 2027
N2 - Reliable monocular depth prediction from endoscopic video frames remains useful for scale-aware scene understanding during minimally invasive procedures and supports applications including surgical 3D reconstruction and intraoperative navigation guidance. However, self-supervised methods based on photometric reconstruction often become unstable in weakly textured tissue regions and near poorly separated tissue interfaces, which may cause blurred depth transitions and distorted local geometry. To mitigate these limitations, we design a boundary-oriented self-supervised endoscopic depth model with hybrid edge-feature enhancement. The proposed model improves boundary detail preservation and local structural consistency by strengthening tissue-boundary-sensitive representations and enhancing cross-scale decoder fusion. Specifically, it consists of a Hybrid Feature Edge Enhancement Module (HFE-EM) and a Multi-Feature Fusion Convolutional Block (MFCB). HFE-EM incorporates complementary edge and texture cues into encoder features to improve boundary-aware representation learning, while MFCB refines decoder-side feature aggregation under global contextual guidance. Experiments on the SCARED dataset, cross-dataset evaluation using Hamlyn, and ablation studies show that our method obtains competitive performance and good generalization ability for depth estimation in endoscopic scenes.
AB - Reliable monocular depth prediction from endoscopic video frames remains useful for scale-aware scene understanding during minimally invasive procedures and supports applications including surgical 3D reconstruction and intraoperative navigation guidance. However, self-supervised methods based on photometric reconstruction often become unstable in weakly textured tissue regions and near poorly separated tissue interfaces, which may cause blurred depth transitions and distorted local geometry. To mitigate these limitations, we design a boundary-oriented self-supervised endoscopic depth model with hybrid edge-feature enhancement. The proposed model improves boundary detail preservation and local structural consistency by strengthening tissue-boundary-sensitive representations and enhancing cross-scale decoder fusion. Specifically, it consists of a Hybrid Feature Edge Enhancement Module (HFE-EM) and a Multi-Feature Fusion Convolutional Block (MFCB). HFE-EM incorporates complementary edge and texture cues into encoder features to improve boundary-aware representation learning, while MFCB refines decoder-side feature aggregation under global contextual guidance. Experiments on the SCARED dataset, cross-dataset evaluation using Hamlyn, and ablation studies show that our method obtains competitive performance and good generalization ability for depth estimation in endoscopic scenes.
KW - Boundary-aware representation
KW - Label-free monocular depth prediction
KW - Surgical endoscopy
KW - Texture enhancement
UR - https://www.scopus.com/pages/publications/105046517275
U2 - 10.1007/978-981-92-3513-1_14
DO - 10.1007/978-981-92-3513-1_14
M3 - Conference contribution
AN - SCOPUS:105046517275
SN - 9789819235124
T3 - Lecture Notes in Computer Science
SP - 164
EP - 175
BT - Advanced Intelligent Computing Technology and Applications - 22nd International Conference on Intelligent Computing, ICIC 2026, Proceedings
A2 - Huang, De-Shuang
A2 - Zhang, Qinhu
A2 - Pan, Yijie
A2 - Zhang, Chuanlei
A2 - Chen, Wei
A2 - Li, Bo
A2 - Bao, Wenzheng
A2 - Premaratne, Prashan
PB - Springer Science and Business Media Deutschland GmbH
T2 - 22nd International Conference on Intelligent Computing, ICIC 2026
Y2 - 22 July 2026 through 26 July 2026
ER -