| 摘 要: 针对水下图像模糊、对比度低及颜色失真等问题,提出基于频域分析与多路特征差分卷积的水下图像细节增强网络。该网络通过高效频率通道模块提取多频谱特征,结合多路特征增强卷积模块聚焦显著梯度信息以增强高频细节表达;并设计双路径注意力融合模块实现跨层级特征拼接与乘性交互,促进多尺度信息融合增强,提升细节恢复能力与色彩还原精度。在 UIEB等数据集上的实验结果显示,该网络在峰值信噪比(26.74dB)、结构相似性(0.94)等指标上优于对比主流方法,尤其在细节恢复与色彩校正上表现优异。 |
| 关键词: 水下图像恢复 差分卷积 频域分析 特征融合 |
|
中图分类号:
文献标识码: A
|
| 基金项目: 国家自然科学基金项目资助(61971272;61471227) |
|
| Underwater Image Detail Restoration Network Based on Frequency-Domain Analysis and Multi-Channel Feature Differential Convolution |
|
LUO Kailun, HE Lifeng, HUANG Yicheng, LIU Yu
|
(School of Electronic Information and Artificial Intelligence, Shaanxi University of Science and Technology, Xi’an 710021, China)
1253194987@qq.com; helifeng@ist.aichi-pu.ac.jp; 2841597913@qq.com; 1871963737@qq.com
|
| Abstract: Aiming at the issues such as blurriness, low contrast, and color distortion in underwater images, an underwater image detail enhancement network based on frequency domain analysis and multi-path feature differential convolution is proposed. The network extracts multi-frequency spectrum features via an efficient frequency channel module, and combines a multi-path feature enhancement convolution module to focus on salient gradient information for enhancing high-frequency detail representation. It also designs a dua-l path attention fusion module to realize cross_x005f_x0002_level feature concatenation and multiplicative interaction, promoting multi-scale information fusion enhancement and improving detail restoration capability and color restoration accuracy. Experimental results on datasets such as UIEB show that the proposed network outperforms mainstream methods in metrics like peak signa-l to-noise ratio (26.74 dB) and structural similarity index (0.94), especially performing excellently in detail restoration and color correction. |
| Keywords: underwater image restoration differential convolution frequency analysis feature fusio |
