雨雾雪场景下的目标检测网络及其可解释研究

doi:10.16182/j.issn1004731x.joss.25-0649

摘要/Abstract

摘要：

为解决极端天气下目标检测性能严重退化的问题，提出了一种基于Kolmogorov-Arnold定理的检测框架KADet。设计了动态Kolmogorov-Arnold变换器，利用可学习非线性激活函数增强对天气退化引入的复杂畸变的建模能力；构建了Kolmogorov-Arnold空间通道网络，集成KAT卷积与空间通道卷积，强化退化场景中目标与背景关系的特征学习；引入改进损失函数约束激活函数优化方向，通过可视化激活函数曲线开展可解释性分析。实验结果表明：KADet在多个合成与真实恶劣天气数据集上的检测精度均优于现有方法，函数曲线可视化揭示了模型针对不同退化类型的差异化响应策略，验证了可学习激活函数在天气退化场景中的有效性与可解释性。

关键词: 目标检测, 视觉退化, Kolmogorov-Arnold网络, 可解释性, 注意力机制

Abstract:

To address the severe degradation of object detection performance under extreme weather conditions, a detection framework based on the Kolmogorov-Arnold theorem, termed KADet, is proposed. A dynamic Kolmogorov-Arnold Transformer is designed, which leverages learnable nonlinear activation functions to enhance the modeling capability for complex distortions introduced by weather degradation. A Kolmogorov-Arnold spatial-channel network is developed by integrating KAT convolution with spatial-channel convolution to strengthen feature learning of relationships between targets and backgrounds in degraded scenes. An improved loss function is introduced to guide the optimization of the activation functions, and interpretability is analyzed through visualization of their curves. Experimental results demonstrate that KADet achieves higher detection accuracy than existing methods on multiple synthetic and real-world adverse-weather datasets. Visualization of the function curves reveals that the model adopts different response strategies for different types of degradation, validating the effectiveness and interpretability of learnable activation functions in weather-degraded scenarios.

Key words: object detection, visual degradation, Kolmogorov-Arnold network, interpretability, attention mechanism

中图分类号:

TP391.4

姜彦吉,崔家瑜,董浩等 . 雨雾雪场景下的目标检测网络及其可解释研究[J]. 系统仿真学报, 2026, 38(6): 1628-1646.

Jiang Yanji,Cui Jiayu,Dong Hao,et al . Object Detection Networks and Their Interpretability in Rain, Fog, and Snow Scenarios[J]. Journal of System Simulation, 2026, 38(6): 1628-1646.

图/表 23

图1

图2

图3

图4

表1

表2

图5

表3

图6

表4

图7

表5

图8

表6

图9

图10

图11

图12

图13

图14

表7

表8

图15

参考文献 38

[1]	郭强, 浦世亮, 张世峰, 等. 适合跨域目标检测的雾霾图像增强[J]. 中国图象图形学报, 2022, 27(5): 1481-1492.
	Guo Qiang, Pu Shiliang, Zhang Shifeng, et al. Cross-domain Object Detection Based Foggy Image Enhancement[J]. Journal of Image and Graphics, 2022, 27(5): 1481-1492.
[2]	Chen Weiting, Huang Zhikai, Tsai Cheng-Che, et al. Learning Multiple Adverse Weather Removal via Two-stage Knowledge Learning and Multi-contrastive Regularization: Toward a Unified Model[C]//2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Piscataway: IEEE, 2022: 17632-17641.
[3]	Wang Yongzhen, Yan Xuefeng, Zhang Kaiwen, et al. TogetherNet: Bridging Image Restoration and Object Detection Together via Dynamic Enhancement Learning[J]. Computer Graphics Forum, 2022, 41(7): 465-476.
[4]	Qin Xu, Wang Zhilin, Bai Yuanchao, et al. FFA-Net: Feature Fusion Attention Network for Single Image Dehazing[J]. Proceedings of the AAAI Conference on Artificial Intelligence, 2020, 34(7): 11908-11915.
[5]	Ren Dongdong, Li Jinbao, Han Meng, et al. Not All Areas Are Equal: A Novel Separation-restoration-fusion Network for Image Raindrop Removal[J]. Computer Graphics Forum, 2020, 39(7): 495-505.
[6]	He Kaiming, Sun Jian, Tang Xiaoou. Single Image Haze Removal Using Dark Channel Prior[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2011, 33(12): 2341-2353.
[7]	Li Boyi, Peng Xiulian, Wang Zhangyang, et al. AOD-net: All-in-one Dehazing Network[C]//2017 IEEE International Conference on Computer Vision (ICCV). Piscataway: IEEE, 2017: 4780-4788.
[8]	Ren Dongwei, Zuo Wangmeng, Hu Qinghua, et al. Progressive Image Deraining Networks: A Better and Simpler Baseline[C]//2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Piscataway: IEEE, 2019: 3932-3941.
[9]	Chen Weiting, Fang Haoyu, Hsieh Cheng-Lin, et al. ALL Snow Removed: Single Image Desnowing Algorithm Using Hierarchical Dual-tree Complex Wavelet Representation and Contradict Channel Loss[C]//2021 IEEE/CVF International Conference on Computer Vision (ICCV). Piscataway: IEEE, 2021: 4176-4185.
[10]	Wu Haiyan, Qu Yanyun, Lin Shaohui, et al. Contrastive Learning for Compact Single Image Dehazing[C]//2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Piscataway: IEEE, 2021: 10546-10555.
[11]	Ren Wenqi, Liu Si, Zhang Hua, et al. Single Image Dehazing via Multi-scale Convolutional Neural Networks[C]//Computer Vision-ECCV 2016. Cham: Springer International Publishing, 2016: 154-169.
[12]	Saito K, Ushiku Yoshitaka, Harada Tatsuya, et al. Strong-weak Distribution Alignment for Adaptive Object Detection[C]//2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Piscataway: IEEE, 2019: 6949-6958.
[13]	Sindagi V A, Oza P, Yasarla R, et al. Prior-based Domain Adaptive Object Detection for Hazy and Rainy Conditions[C]//Computer Vision-ECCV 2020. Cham: Springer International Publishing, 2020: 763-780.
[14]	Zhang Shizhao, Hongya Tuo, Hu Jian, et al. Domain Adaptive YOLO for One-stage Cross-domain Detection[C]//Proceedings of the 13th Asian Conference on Machine Learning. Chia Laguna Resort: PMLR, 2021: 785-797.
[15]	Hnewa M, Radha H. Multiscale Domain Adaptive YOLO for Cross-domain Object Detection[C]//2021 IEEE International Conference on Image Processing (ICIP). Piscataway: IEEE, 2021: 3323-3327.
[16]	Chen Yuhua, Li Wen, Sakaridis Christos, et al. Domain Adaptive Faster R-CNN for Object Detection in the Wild[C]//2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Piscataway: IEEE, 2018: 3339-3348.
[17]	RoyChowdhury A, Chakrabarty P, Singh A, et al. Automatic Adaptation of Object Detectors to New Domains Using Self-training[C]//2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Piscataway: IEEE, 2019: 780-790.
[18]	Rezaeianaran Farzaneh, Shetty Rakshith, Aljundi Rahaf, et al. Seeking Similarities over Differences: Similarity-based Domain Alignment for Adaptive Object Detection[C]//2021 IEEE/CVF International Conference on Computer Vision (ICCV). Piscataway: IEEE, 2021: 9184-9193.
[19]	Wang Xiaofeng, Liu Xiao, Yang Hong, et al. Degradation Modeling for Restoration-enhanced Object Detection in Adverse Weather Scenes[J]. IEEE Transactions on Intelligent Vehicles, 2025, 10(3): 2064-2079.
[20]	Basri Ronen, Galun Meirav, Geifman Amnon, et al. Frequency Bias in Neural Networks for Input of Non-uniform Density[C]//Proceedings of the 37th International Conference on Machine Learning. Chia Laguna Resort: PMLR, 2020: 685-694.
[21]	Liu Ziming, Wang Yixuan, Vaidya S, et al. KAN: Kolmogorov-arnold Networks[EB/OL]. (2025-02-09) [2025-07-01]. .
[22]	包强强, 唐思源, 李擎乾, 等. 改进YOLOv8模型实现多类型肺结节检测[J]. 计算机科学与探索, 2025, 19(2): 429-442.
	Bao Qiangqiang, Tang Siyuan, Li Qingqian, et al. Improved YOLOv8 Model for Multi-type Lung Nodule Detection[J]. Journal of Frontiers of Computer Science and Technology, 2025, 19(2): 429-442.
[23]	Yang Xingyi, Wang Xinchao. Kolmogorov-arnold Transformer[EB/OL]. (2024-09-16) [2025-07-01]. .
[24]	肖志鹏, 何书峰, 田春岐. EmoRepLKNet:一种基于UniRepLKNet的面部情绪识别神经网络[J]. 计算机工程, 2025, 51(11): 54-62.
	Xiao Zhipeng, He Shufeng, Tian Chunqi. EmoRepLKNet: Facial Emotion Recognition Neural Network Based on UniRepLKNet[J]. Computer Engineering, 2025, 51(11): 54-62.
[25]	Lin T Y, Goyal P, Girshick R, et al. Focal Loss for Dense Object Detection[C]//2017 IEEE International Conference on Computer Vision (ICCV). Piscataway: IEEE, 2017: 2999-3007.
[26]	Everingham M, Luc Van Gool, Williams C K I, et al. The Pascal Visual Object Classes (VOC) Challenge[J]. International Journal of Computer Vision, 2010, 88(2): 303-338.
[27]	Sakaridis Christos, Dai Dengxin, Luc van Gool. Semantic Foggy Scene Understanding with Synthetic Data[J]. International Journal of Computer Vision, 2018, 126(9): 973-992.
[28]	Li Boyi, Ren Wenqi, Fu Dengpan, et al. Benchmarking Single-image Dehazing and Beyond[J]. IEEE Transactions on Image Processing, 2019, 28(1): 492-505.
[29]	Ge Zheng, Liu Songtao, Wang Feng, et al. YOLOX: Exceeding YOLO Series in 2021[EB/OL]. (2021-08-06) [2025-07-01]. .
[30]	Carion N, Massa F, Synnaeve G, et al. End-to-end Object Detection with Transformers[C]//Computer Vision-ECCV 2020. Cham: Springer International Publishing, 2020: 213-229.
[31]	Zhao Chuyang, Sun Yifan, Wang Wenhao, et al. MS-DETR: Efficient DETR Training with Mixed Supervision[C]//2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Piscataway: IEEE, 2024: 17027-17036.
[32]	Li Yanghao, Mao Hanzi, Girshick R, et al. Exploring Plain Vision Transformer Backbones for Object Detection[C]//Computer Vision-ECCV 2022. Cham: Springer Nature Switzerland, 2022: 280-296.
[33]	Li Lerenhan, Dong Yunlong, Ren Wenqi, et al. Semi-supervised Image Dehazing[J]. IEEE Transactions on Image Processing, 2020, 29: 2766-2779.
[34]	Wang Xiaofeng, Chen Honggang, Gou Haosong, et al. RestorNet: An Efficient Network for Multiple Degradation Image Restoration[J]. Knowledge-Based Systems, 2023, 282: 111116.
[35]	Huang Shih-Chia, Le Trung-Hieu, Jaw Da-Wei. DSNet: Joint Semantic Learning for Object Detection in Inclement Weather Conditions[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021, 43(8): 2623-2633.
[36]	Liu Wenyu, Ren Gaofeng, Yu Runsheng, et al. Image-adaptive YOLO for Object Detection in Adverse Weather Conditions[J]. Proceedings of the AAAI Conference on Artificial Intelligence, 2022, 36(2): 1792-1800.
[37]	Li Boyun, Liu Xiao, Hu Peng, et al. All-in-one Image Restoration for Unknown Corruption[C]//2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Piscataway: IEEE, 2022: 17431-17441.
[38]	Lin Junhong, Jiang Nanfeng, Zhang Zhentao, et al. LMQFormer: A Laplace-prior-guided Mask Query Transformer for Lightweight Snow Removal[J]. IEEE Transactions on Circuits and Systems for Video Technology, 2023, 33(11): 6225-6235.

数据集	图像	自行车	巴士	汽车	摩托车	人类
VOC-FOG	9 578	836	684	2 453	801	13 519
VOC-Rain	9 578	836	684	2 453	801	13 519
VOC-Snow	9 578	836	684	2 453	801	13 519
VOC-FOG-test	2 129	155	156	857	131	3 527
VOC-Rain-test	2 129	155	156	857	131	3 527
VOC-Snow-test	2 129	155	156	857	131	3 527
FDD	101	17	17	425	9	269
RTTS	4 332	534	1 838	18 413	862	7 950

模型	专攻天气类型	类型	人	自行车	小汽车	摩托车	巴士	平均精度
TogetherNet^[3]	雨、雾、雪	ODAW	83.59	67.60	79.12	71.70	86.27	77.66
FFA-YOLOXs*^[4]	雾	IR+OD	78.30	70.31	69.97	68.80	80.72	73.62
DCP-YOLOXs*^[6]	雾	IR+OD	81.84	70.38	78.63	73.48	84.68	77.80
AOD-YOLOXs*^[7]	雾	IR+OD	67.40	49.19	60.51	55.59	62.07	58.95
RDMNet^[19]	雨、雾、雪	ODAW	83.98	71.18	79.16	73.92	86.19	78.89
YOLOXs^[29]	雨、雾、雪	OD	67.67	83.28	77.75	68.91	81.70	75.86
YOLOXs*	雨、雾、雪	OD	73.09	57.22	69.55	59.83	77.34	67.41
DETR*^[30]	雨、雾、雪	OD	79.74	62.48	70.24	64.16	81.31	71.58
MS-DETR*^[31]	雨、雾、雪	OD	81.41	65.13	78.20	71.08	84.52	76.07
VitDet*^[32]	雨、雾、雪	OD	62.08	30.28	36.70	33.70	54.45	43.44
Semi-YOLOXs*^[33]	雾	IR+OD	81.15	76.94	76.92	72.89	84.88	78.56
RestorNet-YOLOXs*^[34]	雾、雨	IR+OD	78.71	67.15	72.56	71.68	82.36	74.49
DS-Net^[35]	雾	ODAW	72.44	60.47	81.27	53.85	61.43	65.89
IA-YOLO^[36]	雾	ODAW	70.98	61.98	70.98	57.93	61.98	64.77
KADet	雨、雾、雪	ODAW	85.23	69.21	80.86	75.35	88.45	79.82

模型	专攻天气类型	类型	人	自行车	小汽车	摩托车	巴士	mAP
TogetherNet	雨、雾、雪	ODAW	25.39	17.86	56.79	7.14	43.25	30.09
FFA-YOLOXs*	雾	IR+OD	19.18	18.07	50.83	2.38	42.77	26.65
DCP-YOLOXs*	雾	IR+OD	21.57	17.85	55.30	3.57	39.92	27.64
AOD-YOLOXs*	雾	IR+OD	24.54	33.82	56.75	4.76	36.04	31.18
RDMNet	雨、雾、雪	ODAW	25.82	20.31	55.83	2.04	33.60	27.52
YOLOXs	雨、雾、雪	OD	24.37	22.33	55.57	14.29	37.34	30.78
YOLOXs*	雨、雾、雪	OD	21.48	18.84	54.67	1.59	30.33	25.38
DETR*	雨、雾、雪	OD	0.01	0	0.03	0	0	0.01
MS-DETR*	雨、雾、雪	OD	0	0	0.04	0	0	0.01
VitDet*	雨、雾、雪	OD	4.64	0	5.73	0	7.48	3.57
Semi-YOLOXs*	雾	IR+OD	22.39	27.73	56.47	4.76	44.93	31.26
RestorNet-YOLOXs*	雾、雨	IR+OD	23.24	18.81	53.71	2.38	35.70	26.77
DS-Net	雾	ODAW	26.74	20.54	58.16	7.14	36.11	29.74
IA-YOLO	雾	ODAW	16.20	11.76	41.43	4.76	17.55	18.34
KADet	雨、雾、雪	ODAW	27.75	27.01	59.69	14.29	42.20	34.19

模型	专攻天气类型	类型	人	自行车	小汽车	摩托车	巴士	mAP
TogetherNet	雨、雾、雪	ODAW	77.05	42.95	68.80	42.41	28.99	52.04
FFA-YOLOXs*	雾	IR+OD	76.52	48.13	64.31	39.74	23.71	50.48
DCP-YOLOXs*	雾	IR+OD	81.16	51.34	71.13	47.20	31.09	56.38
AOD-YOLOXs*	雾	IR+OD	76.49	43.32	61.03	34.54	22.16	47.51
RDMNet	雨、雾、雪	ODAW	82.25	52.38	71.91	48.98	31.53	57.41
YOLOXs	雨、雾、雪	OD	81.78	56.70	70.23	49.48	31.57	57.95
YOLOXs*	雨、雾、雪	OD	80.28	50.75	68.23	41.89	28.89	54.01
DETR*	雨、雾、雪	OD	60.10	40.22	52.06	31.25	20.03	40.73
MS-DETR*	雨、雾、雪	OD	74.19	19.47	58.07	27.52	19.99	39.85
VitDet*	雨、雾、雪	OD	16.56	1.12	5.24	2.37	0.58	5.17
Semi-YOLOXs*	雾	IR+OD	75.71	46.72	62.74	40.37	24.51	50.01
RestorNet-YOLOXs*	雾、雨	IR+OD	77.48	51.43	60.92	43.12	29.16	52.42
DS-Net	雾	ODAW	68.81	18.02	46.13	15.15	15.44	32.71
IA-YOLO	雾	ODAW	67.25	35.28	41.14	20.97	13.64	35.66
KADet	雨、雾、雪	ODAW	84.42	56.39	76.27	58.14	41.38	63.32

模型	专攻天气类型	类型	人	自行车	小汽车	摩托车	巴士	mAP
TogetherNet	雨、雾、雪	ODAW	83.53	69.83	77.88	74.26	86.41	78.38
RDMNet	雨、雾、雪	ODAW	83.77	72.77	77.86	73.03	86.44	78.77
YOLOXs	雨、雾、雪	OD	80.98	65.88	74.54	70.83	84.35	75.32
YOLOXs*	雨、雾、雪	OD	75.94	63.00	66.92	65.32	73.10	68.86
DETR*	雨、雾、雪	OD	80.24	61.79	70.81	60.51	80.25	70.72
MS-DETR*	雨、雾、雪	OD	81.77	68.93	76.26	69.57	85.78	76.46
VitDet*	雨、雾、雪	OD	61.93	29.97	32.84	36.77	56.16	43.54
RestorNet-YOLOXs*	雾、雨	IR+OD	80.44	68.29	71.68	70.16	82.54	74.62
AirNet-YOLOXs*^[37]	雾、雨	IR+OD	80.06	67.81	70.75	69.87	82.77	74.25
KADet	雨、雾、雪	ODAW	84.32	72.31	78.73	73.43	87.21	79.20