Degradation Type-Aware Image Restoration for Effective Object Detection in Adverse Weather DOI Creative Commons

Xiaochen Huang,

Xiaofeng Wang, Qizhi Teng

и другие.

Sensors, Год журнала: 2024, Номер 24(19), С. 6330 - 6330

Опубликована: Сен. 30, 2024

Despite significant advancements in CNN-based object detection technology, adverse weather conditions can disrupt imaging sensors’ ability to capture clear images, thereby adversely impacting accuracy. Mainstream algorithms for enhance performance through image restoration methods. Nevertheless, the majority of these approaches are designed a specific degradation scenario, making it difficult adapt diverse conditions. To cope with this issue, we put forward type-aware restoration-assisted network, dubbed DTRDNet. It contains an network shared feature encoder (SFE) and decoder, discrimination decoder (DDIR), category predictor (DCP). In training phase, jointly optimize whole framework on mixed dataset, including degraded images clean images. Specifically, type information is incorporated our DDIR avoid interaction between module. Furthermore, DCP makes SFE possess awareness ability, enhancing detector’s adaptability enabling furnish requisite environmental as required. Both be removed according requirement inference stage retain real-time algorithm. Extensive experiments clear, hazy, rainy, snowy demonstrate that DTRDNet outperforms advanced algorithms, achieving average mAP 79.38% across four test sets.

Язык: Английский

A Brief Review on Differentiable Rendering: Recent Advances and Challenges DOI Open Access
Ruicheng Gao, Yue Qi

Electronics, Год журнала: 2024, Номер 13(17), С. 3546 - 3546

Опубликована: Сен. 6, 2024

Differentiable rendering techniques have received significant attention from both industry and academia for novel view synthesis or reconstructing shapes materials one multiple input photographs. These are used to propagate gradients image pixel colors back scene parameters. The obtained can then be in various optimization algorithms reconstruct the representation further propagated into a neural network learn scene’s representations. In this work, we provide brief taxonomy of existing popular differentiable methods, categorizing them based on primary employed: physics-based (PBDR), methods radiance fields (NeRFs), 3D Gaussian splatting (3DGS). Since there already several reviews NeRF-based 3DGS-based but almost zero rendering, place our main focus PBDR and, completeness, only review improvements made NeRF 3DGS survey. Specifically, introductions theories behind all three categories benchmark comparison performance influential works across different aspects, summary current state open research problems. With survey, seek welcome new researchers field offer useful reference key works, inspire future through concluding section.

Язык: Английский

Процитировано

4
Applications of knowledge distillation in remote sensing: A survey DOI
Yassine Himeur, Nour Aburaed, Omar Elharrouss

и другие.

Information Fusion, Год журнала: 2024, Номер unknown, С. 102742 - 102742

Опубликована: Окт. 1, 2024

Язык: Английский

Процитировано

4
Breaking New Ground in Monocular Depth Estimation with Dynamic Iterative Refinement and Scale Consistency DOI Creative Commons
Akmalbek Abdusalomov, Sabina Umirzakova,

Makhkamov Bakhtiyor Shukhratovich

и другие.

Applied Sciences, Год журнала: 2025, Номер 15(2), С. 674 - 674

Опубликована: Янв. 11, 2025

Monocular depth estimation (MDE) is a critical task in computer vision with applications autonomous driving, robotics, and augmented reality. However, predicting from single image poses significant challenges, especially dynamic scenes where moving objects introduce scale ambiguity inaccuracies. In this paper, we propose the Dynamic Iterative Depth Estimation (DI-MDE) framework, which integrates an iterative refinement process novel scale-alignment module to address these issues. Our approach combines elastic bins that adjust dynamically based on uncertainty estimates mechanism ensure consistency between static regions. Leveraging self-supervised learning, DI-MDE does not require ground truth labels, making it scalable applicable real-world environments. Experimental results standard datasets such as SUN RGB-D KITTI demonstrate our method achieves state-of-the-art performance, significantly improving prediction accuracy scenes. This work contributes robust efficient solution challenges of monocular estimation, offering advancements both consistency.

Язык: Английский

Процитировано

0
Performance evaluation of pretrained deep learning architectures for railway passenger ride quality classification DOI
Aliyu Kasimu, Wei Zhou, Qingkai Meng

и другие.

Alexandria Engineering Journal, Год журнала: 2025, Номер 118, С. 194 - 207

Опубликована: Янв. 22, 2025

Язык: Английский

Процитировано

0
Object Extraction-Based Comprehensive Ship Dataset Creation to Improve Ship Fire Detection DOI Creative Commons
Farkhod Akhmedov,

Sanjar Mukhamadiev,

Akmalbek Abdusalomov

и другие.

Fire, Год журнала: 2024, Номер 7(10), С. 345 - 345

Опубликована: Сен. 27, 2024

The detection of ship fires is a critical aspect maritime safety and surveillance, demanding high accuracy in both identification response mechanisms. However, the scarcity fire images poses significant challenge to development training effective machine learning models. This research paper addresses this by exploring advanced data augmentation techniques aimed at enhancing datasets for detection. We have curated dataset comprising (both non-fire) various oceanic images, which serve as target source images. By employing diverse image blending methods, we randomly integrate ships with environments under conditions, such windy, rainy, hazy, cloudy, or open-sky scenarios. approach not only increases quantity but also diversity data, thus improving robustness performance models detecting across different contexts. Furthermore, developed Gradio web interface application that facilitates selective key contribution work related object extraction-based blending. propose basic while applying randomness. Overall, cover eight steps creation. collected 9200 4100 non-fire From augmented 90 13 background achieved 11,440 To test performance, trained Yolo-v8 Yolo-v10 “Fire” “No-fire” In case, precision-recall curve 96.6% (Fire), 98.2% (No-fire), 97.4% mAP score achievement all classes 0.5 rate. model achievement, got 90.3% 93.7 92% comparison, models’ outperforming other Yolo-based SOTA overall scores.

Язык: Английский

Процитировано

1
Degradation Type-Aware Image Restoration for Effective Object Detection in Adverse Weather DOI Creative Commons

Xiaochen Huang,

Xiaofeng Wang, Qizhi Teng

и другие.

Sensors, Год журнала: 2024, Номер 24(19), С. 6330 - 6330

Опубликована: Сен. 30, 2024

Despite significant advancements in CNN-based object detection technology, adverse weather conditions can disrupt imaging sensors’ ability to capture clear images, thereby adversely impacting accuracy. Mainstream algorithms for enhance performance through image restoration methods. Nevertheless, the majority of these approaches are designed a specific degradation scenario, making it difficult adapt diverse conditions. To cope with this issue, we put forward type-aware restoration-assisted network, dubbed DTRDNet. It contains an network shared feature encoder (SFE) and decoder, discrimination decoder (DDIR), category predictor (DCP). In training phase, jointly optimize whole framework on mixed dataset, including degraded images clean images. Specifically, type information is incorporated our DDIR avoid interaction between module. Furthermore, DCP makes SFE possess awareness ability, enhancing detector’s adaptability enabling furnish requisite environmental as required. Both be removed according requirement inference stage retain real-time algorithm. Extensive experiments clear, hazy, rainy, snowy demonstrate that DTRDNet outperforms advanced algorithms, achieving average mAP 79.38% across four test sets.

Язык: Английский

Процитировано

1