Edge-global collaborative enhancement detection for low-contrast infrared sea ice images
-
摘要:
目的： 红外成像因优异的抗雾干扰能力，是恶劣气象下海冰连续监测的核心光学传感模态。针对红外海冰图像中目标与背景灰度差异微弱、边缘模糊的低对比度检测难题，提出一种边缘-全局协同增强检测方法。
方法： 以YOLOv7(You Only Look Once version 7)为基础框架，在浅层ELAN模块中嵌入Sobel轮廓提取模块，通过水平与垂直方向梯度提取及自适应融合，强化海冰边缘的光学梯度特征；在深层ELAN模块中引入大核卷积（LKC）模块，采用主分支瓶颈式大核卷积与短接分支3×3卷积的双分支并行设计，在扩大感受野的同时保留局部细节，将小目标光学斑块与全局辐射背景进行关联建模。
结果： 实验基于多气象场景红外海冰数据集，以原始YOLOv7为基线，构建仅加Sobel、仅加LKC及双模块融合三组对照模型，选取精确率、召回率、mAP@0.5和计算量四项指标进行评价。浓雾条件下，所提方法召回率达0.674，mAP@0.5达0.671，较基线分别提升9.3个百分点和7.7个百分点；全混合气象测试集上召回率达0.63，mAP@0.5达0.627。消融实验表明双模块协同工作时各项指标最优，计算量仅从103.2 GFLOPs微增至103.5 GFLOPs。
结论： 所提方法通过边缘梯度增强与全局背景关联建模的协同，显著提升了低对比度红外海冰目标的检测精度与鲁棒性，基本满足恶劣气象下实时红外海冰监测的需求。
Abstract:ObjectiveInfrared imaging serves as a core sensing modality for sea ice monitoring under severe weather conditions, owing to its excellent fog-penetration capability that supports continuous observation. However, low contrast and blurred edges remain key challenges in infrared sea ice imagery.
MethodThis paper proposes an edge-global collaborative enhancement detection method built upon the YOLOv7 (You Only Look Once version 7) framework. A Sobel contour extraction module is embedded in the shallow ELAN blocks, which extracts horizontal and vertical gradients and performs adaptive fusion to enhance the edge gradient features of sea ice boundaries. A Large Kernel Convolution (LKC) module is then added to the deep ELAN blocks, adopting a dual-branch parallel complementary design: the main bottleneck branch expands the network's receptive field, while the short-connected 3×3 convolution branch preserves fine local details of targets and correlates local image patches with the global radiative background.
ResultExperiments are conducted on a multi-weather infrared sea ice dataset, with the original YOLOv7 set as the baseline for comparison. Three control models are constructed: Sobel-only, LKC-only, and dual-module. Four metrics are adopted for evaluation: precision, recall, mAP@0.5, and computational cost. Under dense fog conditions, the proposed method achieves a recall of 0.674 and an mAP@0.5 of 0.671, corresponding to improvements of 9.3 and 7.7 percentage points over the baseline, respectively. On the mixed-weather test set, recall reaches 0.63 and mAP@0.5 reaches 0.627. Ablation experiments verify that the dual-module model delivers the best performance across all metrics, with only a slight increase in computational cost from 103.2 GFLOPs to 103.5 GFLOPs.
ConclusionBy integrating edge enhancement and global background modeling, the proposed method significantly improves the detection accuracy of low-contrast sea ice targets and enhances model robustness in harsh weather scenarios, and can effectively meet the requirements of real-time sea ice monitoring under adverse weather conditions.
-
图 1 红外子数据集部分气象图示
Figure 1. Selected Meteorological Scenario Diagrams from the Infrared Sub-dataset
图 5 Sobel算子和LKC模块在ELAN模块中的嵌入位置
Figure 5. Embedding Positions of the Sobel Operator and LKC Module in the ELAN Module
图 6 数据集中不同气象的图像示例
注：(a):晴天 (b):轻雾 (c):浓雾 (d):雪天 (e):雨天
Figure 6. Image Samples Under Different Weather Conditions in the Dataset
图 7 基线网络和改进网络的检测结果
注：第一行为基线网络检测结果，第二行为改进网络结果
Figure 7. Detection Result Comparison Between the Baseline Network and the Improved Network
表 6 单一模块在轻雾场景下的测试数据
Table 6. Performance of Individual Modules on the Light Fog Test Set
Weather Fog2 Model Comput. P R mAP@0.5 Baseline 103.2 0.928 0.605 0.618 Baseline+Sobel 103.2 0.879 0.674 0.641 Baseline+LKC 103.5 0.951 0.605 0.646 Baseline+Sobel+LKC 103.5 0.886 0.721 0.705 
下载: 导出CSV
表 7 单一模块在浓雾场景下的测试数据
Table 7. Performance of Individual Modules on the Dense Fog Test Set
Weather Fog5 Model Comput. P R mAP@0.5 Baseline 103.2 0.893 0.581 0.594 Baseline+Sobel 103.2 0.999 0.628 0.631 Baseline+LKC 103.5 0.812 0.605 0.600 Baseline+Sobel+LKC 103.5 0.967 0.674 0.671
下载: 导出CSV
表 1 混合气象以及晴天测试集测试结果
Table 1. Test Results on Mixed Weather and Sunny Test Sets
Weather All Clear Model Comput. P R mAP@0.5 P R mAP@0.5 Baseline 103.2 0.788 0.551 0.558 0.971 0.439 0.497 Baseline+
Sobel+LKC103.5 0.830 0.630 0.627 0.864 0.507 0.533
下载: 导出CSV
表 2 轻雾及浓雾测试集测试结果
Table 2. Test Results on Light Fog and Dense Fog Test Sets
Weather Fog2 Fog5 Model Comput. P R mAP@0.5 P R mAP@0.5 Baseline 103.2 0.928 0.605 0.618 0.893 0.581 0.594 Baseline+
Sobel+LKC103.5 0.886 0.721 0.705 0.967 0.674 0.671
下载: 导出CSV
表 3 不同YOLO系列检测器在混合气象测试集（All）上的检测指标对比
Table 3. Comparison of Detection Metrics for Different YOLO-Series Detectors on the Mixed-Weather Test Set
Model P R mAP@0.5 YOLOv5 0.808 0.545 0.636 YOLOv8 0.838 0.623 0.701 YOLOv11 0.776 0.584 0.652 YOLOv7(Baseline) 0.788 0.551 0.558 Ours 0.830 0.630 0.627
下载: 导出CSV
表 4 单一模块在混合场景下的测试数据
Table 4. Performance of Individual Modules on the Mixed-Weather Test Set
Weather All Model Comput. P R mAP@0.5 Baseline 103.2 0.788 0.551 0.558 Baseline+Sobel 103.2 0.834 0.601 0.592 Baseline+LKC 103.5 0.770 0.568 0.565 Baseline+Sobel+LKC 103.5 0.830 0.630 0.627
下载: 导出CSV
表 5 单一模块在晴天场景下的测试数据
Table 5. Performance of Individual Modules on the Clear Weather Test Set
Weather Clear Model Comput. P R mAP@0.5 Baseline 103.2 0.971 0.439 0.497 Baseline+Sobel 103.2 0.818 0.480 0.532 Baseline+LKC 103.5 0.921 0.451 0.510 Baseline+Sobel+LKC 103.5 0.864 0.507 0.533
下载: 导出CSV
-
[1] 李明慧. 基于深度学习的SAR影像海冰分类研究[D]. 上海: 上海海洋大学, 2019.LI M H. Sea ice classification based on deep learning with SAR imagery[D]. Shanghai: Shanghai Ocean University, 2019. (in Chinese). [2] BIJELIC M, GRUBER T, MANNAN F, et al. Seeing through fog without seeing fog: deep multimodal sensor fusion in unseen adverse weather[C]. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, IEEE, 2020: 11679-11689. [3] 陈志昆, 魏立新, 李志强, 等. 2017年夏季北冰洋浮冰区海雾特征分析[J]. 海洋预报, 2019, 36(2): 77-87.CHEN ZH K, WEI L X, LI ZH Q, et al. Sea fog characteristics over the Arctic pack ice in summer 2017[J]. Marine Forecasts, 2019, 36(2): 77-87. (in Chinese). [4] WANG CH X, WU H J, JIN ZH. FourLLIE: boosting low-light image enhancement by Fourier frequency information[C]. Proceedings of the 31st ACM International Conference on Multimedia, Association for Computing Machinery, 2023: 7459-7469. [5] KIM Y T. Contrast enhancement using brightness preserving bi-histogram equalization[J]. IEEE Transactions on Consumer Electronics, 1997, 43(1): 1-8. doi: 10.1109/30.580378 [6] Ancuti C, Ancuti C O, Haber T, et al. Enhancing underwater images and videos by fusion[C]. Proceedings of the 2012 IEEE Conference on Computer Vision and Pattern Recognition, IEEE, 2012: 81-88. [7] REN X Y, JIAO B Y, PENG ZH M, et al. MSFFNet: a multilevel sparse feature fusion network for infrared dim small target detection[J]. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2025, 18: 147-159. doi: 10.1109/JSTARS.2024.3488698 [8] 蔚婧, 张磊, 丁美琪, 等. 空频域特征融合的红外弱小目标检测网络[J]. 西安电子科技大学学报, 2025, 52(6): 32-44.YU J, ZHANG L, DING M Q, et al. Infrared dim and small target detection network with spatial-frequency feature fusion[J]. Journal of Xidian University, 2025, 52(6): 32-44. (in Chinese). [9] 王伯霄, 宋延嵩, 董小娜. 难点注意力感知红外小目标检测网络[J]. 中国光学(中英文), 2024, 17(3): 538-547.WANG B X, SONG Y S, DONG X N. Indistinguishable points attention-aware network for infrared small object detection[J]. Chinese Optics, 2024, 17(3): 538-547. (in Chinese). [10] 周卓, 檀帅兵, 白昆, 等. 多尺度空间与形态学特征的红外小目标检测方法[J]. 弹箭与制导学报, 2026, 46(3): 291-304.ZHOU ZH, TAN SH B, BAI K, et al. Infrared small target detection based on multi-scale spatial loss and morphological features[J]. Journal of Projectiles, Rockets, Missiles and Guidance, 2026, 46(3): 291-304. (in Chinese). [11] 陈科, 蒋行国, 林国军, 等. 红外小目标检测的轻量级ERMR-DETR算法[J]. 激光与红外, 2025, 55(12): 1950-1957.CHEN K, JIANG X G, LIN G J, et al. Lightweight ERMR-DETR algorithm for infrared small target detection[J]. Laser & Infrared, 2025, 55(12): 1950-1957. (in Chinese). [12] 赵佳乐, 娄树理, 林超. 多光谱融合的红外舰船目标轻量化检测[J]. 光学 精密工程, 2025, 33(8): 1327-1338.ZHAO J L, LOU SH L, LIN CH. A lightweight detection of multi-spectral infrared ship targe[J]. Optics and Precision Engineering, 2025, 33(8): 1327-1338. (in Chinese). [13] TANG Y, XU T F, QIN H L, et al. IRSTD-YOLO: an improved YOLO framework for infrared small target detection[J]. IEEE Geoscience and Remote Sensing Letters, 2025, 22: 7001405. doi: 10.1109/lgrs.2025.3562096 [14] 薛驰, 陈小梅, 李海彤. 融合背景估计与相对局部对比度的天基短波红外弱小目标检测[J]. 光学 精密工程, 2026, 34(3): 450-465.XUE CH, CHEN X M, LI H T. SWIR weak targets detection on space-based platform integrating background estimation and relative local contrast[J]. Optics and Precision Engineering, 2026, 34(3): 450-465. (in Chinese). [15] 赵阳, 杨文贵, 高翠云. 基于全局双组注意力的红外与可见光图像融合[J]. 液晶与显示, 2025, 40(12): 1840-1852.ZHAO Y, YANG W G, GAO C Y. Infrared and visible image fusion based on global dual-group attention[J]. Chinese Journal of Liquid Crystals and Displays, 2025, 40(12): 1840-1852. (in Chinese). [16] CHEN Y SH, WANG B R, GUO X Y, et al. DEYOLO: dual-feature-enhancement YOLO for cross-modality object detection[C]. 27th International Conference on Pattern Recognition, Springer, 2024: 236-252. [17] 刘杰, 安博文. 海面红外小目标检测算法研究[J]. 红外技术, 2015, 37(1): 16-19.LIU J, AN B W. Research on the detection algorithm for infrared small target on the sea[J]. Infrared Technology, 2015, 37(1): 16-19. (in Chinese). -
下载:
计量
- 文章访问数: 5
- HTML全文浏览量: 1
- PDF下载量: 1
- 被引次数: 0
