ADPCNet: Adaptive Deformable Peripheral Convolution for Efficient Image Dehazing

Area of Science:

• Computer Vision
• Image Processing
• Deep Learning

Background:

• Single-image dehazing is challenging due to spatially varying degradation.
• Existing large-context models often suffer from high computational costs or loss of fine details.

Purpose of the Study:

• To propose an efficient and effective single-image dehazing method.
• To address limitations of current models in handling complex haze patterns and preserving image details.

Main Methods:

• Introduced the Adaptive Deformable Peripheral Convolution Network (ADPCNet), a compact encoder-decoder architecture.
• Employed conditional adaptive sharing for context modeling, deformable sampling for aggregation, frequency-guided modulation for detail compensation, and dynamic multi-branch fusion.
• Separated haze estimation, structure alignment, and detail recovery into an efficient operator stack.

Main Results:

• ADPCNet achieved competitive performance on RESIDE, Dense-Haze, and NH-Haze datasets.
• Demonstrated strong results on SOTS-Indoor (40.89 dB/0.997) and SOTS-Outdoor (37.80 dB/0.996).
• Achieved a favorable quality-efficiency trade-off with 7.25 M parameters and 33.62 G FLOPs.

Conclusions:

• ADPCNet offers an effective solution for single-image dehazing.
• The proposed modules contribute significantly to improved performance and efficiency.
• The network maintains a good balance between image quality and computational cost.