Related Experiment Videos
MIP-Net: Multi-scale inverted pyramid network with local adaptive contrast for underwater object detection
Peixin Qu1, Yiteng Guo1, Guohou Li1
1School of Information Engineering, Henan Institute of Science and Technology, Xinxiang, 453003, China.
Summary
This study introduces the Multi-Scale Inverted Pyramid Network (MIP-Net) to improve underwater object detection (UOD) by enhancing feature contrast and separating foreground from background noise. MIP-Net achieves superior accuracy on benchmarks, demonstrating its effectiveness in challenging underwater environments.
Area of Science:
- Computer Vision
- Artificial Intelligence
- Robotics
Background:
- Underwater object detection (UOD) faces challenges due to visual degradation like noise, blur, and small object suppression.
- Existing methods struggle with effective feature fusion and handling complex underwater environments.
Purpose of the Study:
- To propose a novel network, the Multi-Scale Inverted Pyramid Network (MIP-Net), for robust underwater object detection.
- To address limitations in current UOD techniques by introducing innovative feature enhancement and separation modules.
Main Methods:
- Developed the Local Adaptive Contrast (LAC) module for layer-by-layer intra-feature contrast calibration, preventing detail dilution.
- Introduced the Multi-Scale Inverted Feature Pyramid Network (MSIFPN) with dual-pyramid bidirectional flow and foreground-background separation.
- Implemented a strict mathematical strategy for isolating target semantics from water impurities.
Main Results:
- MIP-Net achieved an Average Precision (AP) of 70.1% on the DUO benchmark, outperforming state-of-the-art methods.
- Demonstrated strong cross-domain generalization with an AP of 45.6% on the terrestrial COCO dataset.
- Achieved a competitive balance between computational efficiency (Params/FLOPs) and detection accuracy.
Conclusions:
- MIP-Net effectively tackles visual degradation challenges in UOD through its novel LAC and MSIFPN components.
- The proposed framework shows significant improvements in accuracy and generalization capabilities for underwater object detection.
- The public availability of the code facilitates further research and development in this domain.