Core Concepts
Proposing HCF-Net to enhance infrared small object detection performance through practical modules.
Abstract
Infrared small object detection is crucial but challenging due to the diminutive size of objects and complex backgrounds in images. The proposed HCF-Net introduces modules like PPA, DASI, and MDCR to address these challenges effectively. PPA uses multi-branch feature extraction, DASI enables adaptive channel selection, and MDCR captures spatial features through multiple convolutional layers. The model is evaluated on the SIRST dataset, outperforming traditional methods with significant advantages. The network architecture includes an encoder-decoder structure with key modules strategically placed for improved performance.
Stats
Extensive experimental results on the SIRST infrared single-frame image dataset show that the proposed HCF-Net performs well.
The computational cost of HCF-Net is 93.16 GMac with 15.29 million parameters.
Quotes
"Our contributions in this paper can be summarized as modeling infrared small object detection as a semantic segmentation problem and proposing HCF-Net."
"The proposed method achieves outstanding performance on the SIRST dataset, significantly surpassing other methods."
"HCF-Net incorporates multiple practical modules that significantly enhance small object detection performance."