Ijraset Journal For Research in Applied Science and Engineering Technology
Authors: Dr. Aparna Hambarde, Aditya Kotame, Vrushali Khade, Samruddhi Latore, Vaishnavi Jedhe
DOI Link: https://doi.org/10.22214/ijraset.2025.66525
Certificate: View Certificate
This paper presents an AI-ML-based defogging and smoke removal algorithm that uses a hybrid model combined with a convolutional neural network (CNN) and Transformer. The proposed model uses the U-Net backbone to extract local detailed information, uses the FFA-Net maintenance mechanism to update regions with different burning conditions, and uses the Dehamer transformer encoder to capture and manage the comprehensive information of remote locations in the image. This combination works well in a wide range of climate and smoke conditions, producing high-quality images with accurate colors. The model is evaluated on synthetic and real data, yielding results that outperform today\'s traditional methods and deep learning models. Its effective capabilities make it suitable for immediate use, making it possible to eliminate and eliminate smoke technology in difficult situations.
I. INTRODUCTION
Haze and smoke, caused by atmospheric conditions such as fog, mist, and pollution, present significant challenges in capturing clear and accurate images. These environmental factors degrade visibility and introduce color distortions, which negatively impact various computer vision applications, including autonomous driving, surveillance, and disaster management. In real-world scenarios, haze and smoke often exhibit non-uniform distributions, making it difficult for traditional image restoration methods to effectively enhance image clarity. Consequently, there is a growing demand for advanced techniques capable of addressing these complex and dynamic conditions.
Traditional approaches, such as the Dark Channel Prior (DCP) introduced by He et al. (2009), have been widely utilized to estimate haze density and improve visibility. While effective in some cases, these methods are often limited when addressing bright regions like skies or reflective surfaces, frequently resulting in color distortions and halo artifacts. Moreover, DCP-based algorithms struggle with non-uniform haze, where the varying density across the image leads to inconsistent dehazing outcomes.
Recent advancements in deep learning have paved the way for more sophisticated methods in haze and smoke removal. Convolutional Neural Networks (CNNs) have demonstrated their effectiveness in extracting spatial features from images, while attention mechanisms and transformer models have shown great promise in capturing global context and modeling long-range dependencies. Building on these innovations, this paper introduces an AI-ML-based intelligent dehazing and desmoking algorithm that addresses the limitations of traditional methods by leveraging a hybrid approach combining CNNs, attention mechanisms, and transformers.
The proposed algorithm employs a U-Net backbone for efficient local feature extraction, enabling the model to capture intricate haze or smoke details across different regions of the image. Additionally, it integrates the Feature Fusion Attention Network (FFA-Net) module, which dynamically prioritizes regions with varying haze intensities, effectively managing non-uniform haze.
This paper offers a comprehensive analysis of the proposed method, detailing its architectural components and evaluating its performance on both synthetic and real-world datasets. Experimental results demonstrate that the AI-ML-based solution outperforms traditional dehazing techniques and achieves superior results compared to state-of-the-art deep learning models. Furthermore, its computational efficiency makes it suitable for real-time applications, positioning it as a valuable tool for critical computer vision tasks.
The remainder of this paper is structured as follows: Section 2 reviews related research on dehazing and desmoking. Section 3 outlines the methodology underlying the proposed model, including its CNN backbone, attention mechanisms, and transformer encoder. Section 4 describes the experimental setup and evaluation metrics, with the results and discussion presented in Section 5. Finally, Section 6 concludes the paper and discusses potential directions for future research.
II. RELATED WORK
This section explores significant research in the field of image dehazing and desmoking, with a particular focus on comparing conventional methods with modern deep learning techniques. The analysis provided compares the performance of each approach under varying haze and smoke intensities, illustrated using a sample image.
A. Advances in Dehazing Algorithms:
B. Key differences become evident:
The results from the sample image clearly indicate that deep learning models, particularly FFA-Net and Dehamer, surpass traditional techniques in terms of image quality, especially in environments with non-uniform haze and smoke distributions.
Figure 1: Example of dehazing results from Non-homogeneous hazy images of varying thickness. Source: Vinay P. et al. [1] DOI: 10.1109/WACVW58289.2023.00061
ISSN: 2690-621X
This section presents a comparative analysis of traditional and advanced deep learning-based dehazing methods, focusing on their effectiveness in addressing varying haze and smoke intensities. Sample images processed by multiple algorithms highlight the improvements modern techniques provide over conventional approaches.
This analysis underscores the evolution of dehazing techniques from traditional methods to advanced deep learning models, demonstrating the significant performance improvements offered by hybrid architectures.
III. METHODOLOGY
This section outlines the methodology used to develop our AI/ML-based intelligent de-smoking and de-hazing algorithm. The proposed solution incorporates a combination of advanced techniques such as Convolutional Neural Networks (CNN), attention mechanisms, and transformer encoders to enhance image quality in the presence of haze and smoke.
1) Data Collection and Preprocessing: The first step involves gathering a varied dataset of hazy and smoky images sourced from publicly available datasets and synthetic images generated using haze simulation techniques. Each image undergoes several preprocessing stages to standardize the input, which include:
2) Architecture Design: The system’s architecture comprises three key components:
CNN Backbone (e.g., U-Net):
The first component of the system utilizes a Convolutional Neural Network (CNN) backbone, specifically U-Net, which is renowned for its efficacy in image segmentation and restoration tasks.
Its encoder-decoder structure allows for effective extraction of multi-scale features, enabling the model to retain fine details in the final output. This makes U-Net particularly suitable for tasks requiring precise spatial hierarchies.
Attention Mechanisms (e.g., FFA-Net)
To address the challenge of non-uniform haze distributions, attention mechanisms are incorporated into the system. The FFA-Net (Feature Fusion Attention Network) is employed to dynamically focus on different regions of the image, allowing the model to prioritize critical features while minimizing irrelevant background noise. The attention mechanism assigns different weights to image regions, significantly enhancing the quality of the dehazing and desmoking process.
Transformer Encoder (e.g., Dehamer)
The third component of the architecture is the transformer encoder, designed to capture global context and long-range dependencies across the image. The Dehamer architecture is employed here, enabling the model to establish relationships between distant pixels. This capability is essential for the accurate removal of haze and smoke, as it helps the model make informed predictions even when haze intensity varies across the image.
3) Training Strategy
The model is trained using a supervised learning approach with paired hazy/smoky and clear images. The following aspects guide the training process:
4) Evaluation Metrics
To evaluate the performance of the proposed model, we use a combination of quantitative and qualitative metrics:
5) Implementation: The model is implemented using Python, with deep learning frameworks such as TensorFlow and PyTorch for training and deployment. The system is tested on a variety of hazy and smoky images to evaluate its generalization capability and robustness across different scenarios.
6) Post-Processing: Once the model generates the output, a post-processing stage is applied to further enhance the results. Techniques such as contrast enhancement and color correction are used to refine the images, ensuring that the final output is visually appealing and suitable for real-world applications.
The Following flowchart illustrates the systematic process of the AI-based Intelligent De-Hazing and De-Smoking Algorithm, detailing the sequential steps from image input to final output. The process begins with the input stage, where a potentially hazy or smoky image is provided to the system. This input undergoes a pre-processing stage to standardize and enhance it through resizing, normalization, and noise reduction. A decision point verifies whether pre-processing is complete, ensuring image quality before progressing to the next stage. If not, the system reprocesses the image to achieve the required standard.
Following pre-processing, a Convolutional Neural Network (CNN) extracts essential features, such as textures, patterns, and edges, from the image. This feature extraction is critical for understanding the image's spatial structure and serves as the foundation for further processing. An attention mechanism is then applied to focus on the most significant regions within the image, emphasizing areas with higher obstruction while suppressing less relevant features. This step is particularly effective in managing non-uniform haze or smoke.
Fig .2. Flowchart
Subsequently, a transformer encoder captures long-range dependencies and global spatial relationships within the image, leveraging its ability to process contextual information comprehensively. This stage refines the processed data, ensuring that the algorithm can effectively handle complex haze and smoke patterns. Finally, the processed features are used to reconstruct a clear, de-hazed, or de-smoked image. The system produces an output that retains natural color, detail, and structure, ensuring usability in applications such as surveillance, autonomous driving, or environmental monitoring. The process concludes, ready for further iterations If needed.
This research paper presents a comprehensive approach to addressing the challenges posed by haze and smoke in outdoor imagery through an AI-ML-Based Intelligent De-Hazing and De-Smoking Algorithm. Leveraging advancements in deep learning, the proposed system combines the strengths of Convolutional Neural Networks (CNNs), attention mechanisms, and transformers within a hybrid architecture. The integration of U-Net for local feature extraction, FFA-Net for region-specific focus, and Dehamer for global context modeling enables the system to effectively manage non-uniform haze and smoke conditions. Traditional dehazing techniques, such as the Dark Channel Prior (DCP) and Fast Visibility Restoration (FVR), have provided foundational methods but suffer from limitations, including artifacts, color distortions, and poor handling of bright or non-homogeneous haze regions. While these methods are computationally efficient, they lack the robustness required for complex real-world scenarios. Conversely, recent deep learning advancements, including attention mechanisms and transformer-based models, address these challenges by improving clarity, color fidelity, and structural integrity in processed images. This research highlights the transformative potential of combining CNNs, attention mechanisms, and transformers, setting new benchmarks in dehazing and desmoking performance. Future work could explore optimizing computational requirements for transformer components, integrating real-time dynamic feedback mechanisms, and extending the system to handle other atmospheric disturbances, such as rain or dust. This study thus represents a significant advancement in image restoration, offering a reliable solution for critical computer vision tasks in challenging environments.
[1] Vinay, P., Abhisheka, K. S., Shetty, L., Kushal, T. M., & Shylaja, S. S. (2023). “Non Homogeneous Realistic Single Image Dehazing”. Proceedings - 2023 IEEE/CVF Winter Conference on Applications of Computer Vision Workshops, WACVW 2023. DOI: 10.1109/WACVW58289.2023.00061. ISSN: 2690- 621X. Published in 2023. [2] Li, S., Cheng, Y., & Dai, Y. (2012). “Progressive Hybrid-Modulated Network for Single Image Deraining”. In Proceedings 2012 IEEE International Conference on Computer Science and Automation Engineering: May 25-27, 2012, Zhangjiajie, China. IEEE. ISBN: 978-1-4673-2144-9. Published in 2012. [3] Zhang, Y., Gao, K., Wang, J., Zhang, X., Wang, H., Hua, Z., & Wu, Q. (2021). “Single-Image Dehazing Using Extreme Reflectance Channel Prior”. IEEE Access, 9, 87826–87838. DOI: 10.1109/ACCESS.2021.3090202. ISSN: 2169-3536. Published on July 16, 2021. [4] Zhang, Z., & Xie, Y. (2020). “Deep Image Dehazing Using Generative Adversarial Networks”. IEEE International Conference on Wireless Communications Signal Processing and Networking (WiSPNET), 2020, IEEE Transactions on Circuits and Systems for Video Technology, 30(8), 2610-2623. DOI: 10.1109/TCSVT.2020.2979461. ISSN: 1051-8215. Published in 2020. [5] Cai, B., Xu, X., & Jia, J. (2016). “DehazeNet: An End-to-End System for Single Image Haze Removal”. IEEE Transactions on Image Processing, 25(11), 4987-4998. DOI: 10.1109/TIP.2016.2599057. ISSN: 1057-7149. Published on November 2016. [6] He, K., Sun, J., & Tang, X. (2010). \"Single Image Haze Removal Using Dark Channel Prior\". IEEE Transactions on Pattern Analysis and Machine Intelligence, 33(12), 2341–2353. DOI: 10.1109/TPAMI.2010.168. ISSN: 0162-8828. Published in December 2010. [7] Berman, D., Treibitz, T., & Avidan, S. (2016). \"Non-local Image Dehazing\". IEEE Transactions on Pattern Analysis and Machine Intelligence, 38(12), 2419–2432. DOI: 10.1109/TPAMI.2016.2544710. ISSN: 0162-8828. Published in December 2016. [8] Fattal, R. (2008). \"Single Image Dehazing\". ACM Transactions on Graphics, 27(3), 1–9. DOI: 10.1145/1360612.1360673. ISSN: 0730-0301. Published in July 2008. [9] Zhang, L., & Wang, X. (2019). \"Enhancing the Dehazing Network for Low-light Image\". International Journal of Computer Vision, 128(1), 79–95. DOI: 10.1007/s11263-019-01234-3. ISSN: 0920-5691. Published in January 2019. [10] Li, H., & Tan, R. T. (2018). \"A Deep Network for Image Dehazing\". IEEE Transactions on Image Processing, 27(10), 5074–5087. DOI: 10.1109/TIP.2018.2822830. ISSN: 1057-7149. Published in October 2018. [11] Ren, W., Liu, L., & Xu, Y. (2019). \"Learning to Remove Haze in Real-World Images\". IEEE Transactions on Image Processing, 28(10), 5075–5088. DOI: 10.1109/TIP.2019.2907280. ISSN: 1057- 7149. Published in October 2019. [12] Luo, Z., Xie, J., & Yu, W. (2018). \"Real-time Single Image Dehazing Using Convolutional Neural Networks\". Journal of Visual Communication and Image Representation, 46, 242-251. DOI: 10.1016/j.jvcir.2018.06.004. ISSN: 1047-3203. Published in June 2018. [13] Chen, Y., Yu, Z., & Feng, J. (2020). \"A Fast Dehazing Algorithm Using Non-Local Mean and Dark Channel Prior\". Journal of Computer Science and Technology, 35(6), 1320–1333. DOI: 10.1007/s11390- 020-0201-7. ISSN: 1673-4784. Published in November 2020. [14] Yang, X., Li, X., & Li, Z. (2020). \"Image Dehazing Using Deep Generative Networks\". IEEE Transactions on Image Processing, 29, 2901–2916. DOI: 10.1109/TIP.2020.2972892. ISSN: 1057-7149. Published in February 2020. [15] Dong, X., & Yang, X. (2021). \"Learning to Dehaze with Hybrid Loss Function\". Journal of Signal Processing Systems, 93(3), 373-384. DOI: 10.1007/s11265-021-01591-3. ISSN: 1936-1999. Published in March 2021. [16] Wang, S., Zhang, L., & Li, Q. (2018). \"Multi-Scale Image Dehazing with GANs\". IEEE Access, 6, 11328-11336. DOI: 10.1109/ACCESS.2018.2808440. ISSN: 2169-3536. Published on March 6, 2018. [17] Hu, X., Liu, S., & Zhang, J. (2017). \"A Novel Image Dehazing Algorithm Using CNNs\". Journal of Visual Communication and Image Representation, 46, 242-251. DOI: 10.1016/j.jvcir.2017.01.001. ISSN: 1047-3203. Published in January 2017. [18] Hu, Y., & Zhang, Q. (2020). \"A Hybrid Dehazing Network for Low-Level Vision Tasks\". IEEE Transactions on Image Processing, 29, 1992-2004. DOI: 10.1109/TIP.2020.2973414. ISSN: 1057-7149. Published in May 2020. [19] Wu, X., Li, L., & Zhang, Z. (2021). \"Single Image Dehazing with Contextual Attention\". IEEE Transactions on Image Processing, 30, 3898-3910. DOI: 10.1109/TIP.2021.3058257. ISSN: 1057-7149. Published in April 2021. [20] Li, X., & Sun, C. (2019). \"Real-Time Image Dehazing Using Dual-Path Network\". IEEE Transactions on Image Processing, 28(5), 2325-2335. DOI: 10.1109/TIP.2019.2903174. ISSN: 1057-7149. Published in May 2019.
Copyright © 2025 Dr. Aparna Hambarde, Aditya Kotame, Vrushali Khade, Samruddhi Latore, Vaishnavi Jedhe. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
Paper Id : IJRASET66525
Publish Date : 2025-01-14
ISSN : 2321-9653
Publisher Name : IJRASET
DOI Link : Click Here