Lens Flare Artifact Removal: Bridging Traditional Methods with Deep Learning Models (Part-4)

DATASETS IN FLARE REMOVAL 

The dataset by comprises a total of 5,001 images, out of which 3,000 are synthetic and 2,001 are real-world captures. The dataset includes a basic variety of light sources but does not explicitly cover reflective flares, streaks, or glare components. It was one of the earlier contributions used primarily for daytime flare removal training tasks. In contrast, the Flare7K dataset comprises 7,000 synthetic images and was developed to support nighttime flare removal. This dataset is remarkably comprehensive, including a wide range of flare artifacts such as reflective flare, streaks, and glare. It offers greater variability with 25 different light source positions and 10 types of flare patterns, making it well-suited for training deep learning models to handle diverse and complex flare conditions. With data augmentation and testing on PSNR and SSIM basis, their approach showed clear improvements over traditional methods.

According to Table VII. Although smaller with only 962 images, the Flare-R dataset comprises entirely real-world images. It includes 962 unique light source positions and nine flare types, making it valuable for evaluating real-world generalization. However, it lacks examples of reflective flares, streaks, and glare, which limits its use for models requiring exposure to these specific artifacts during training. This comparative analysis underscores the trade-off between realism and variety: while synthetic datasets like Flare7K provide controlled diversity, real datasets like Flare-R better represent actual image acquisition conditions. An ideal flare removal model benefits from exposure to both synthetic diversity and real-world authenticity during training and evaluation.

The tables below compare datasets, methods, and performance metrics such as PSNR, SSIM, and LPIPS.

TABLE V: Comparison of Flare Removal Datasets 

Table VI presents a comparative summary of three prominent datasets commonly used in lens flare removal research: the dataset introduced by Semi-Synthetic Dataset, Flare7K, and Flare-R. These datasets vary in terms of their size, composition (synthetic vs. real images), and the types of flare artifacts included. (N: No, Y: Yes) Semi-Synthetic Dataset, Flare7K, and Flare-R were selected due to their prevalence in literature and coverage of synthetic/real-world flare types. Other datasets lacked public availability or standardized benchmarks.**** 

To determine how well flare removal models work, we look at both objective performance scores and how their architectures compare. In this section, we'll dig into how recent flare removal methods perform, primarily focusing on MFDNet. We will see how it stacks up against other well-known options like Uformer, MPRNet, and Restormer. Table V highlights key attributes such as dataset type, size, flare types covered, and real vs. synthetic composition. While synthetic datasets like Flare7K provide large-scale training data, real or mixed datasets offer better generalization to real-world flare conditions. Limitations include a lack of real diversity, small size, or incomplete documentation. We test these models on both synthetic data and real-world images to see how well they generalize and stay reliable across different scenarios.**** 

When we're trying to see how well flare removal algorithms perform, as shown in Table VIII, we usually look at three main evaluation metrics that everyone agrees on:**** 

Models like Uformer and Restormer, both based on transformer tech, get pretty good PSNR and SSIM scores, especially on synthetic data. Still, they’re a little behind when it comes to making the images look totally natural, as shown by their higher LPIPS scores. According to Figure 6, Higher SSIM and lower LPIPS reflect better visual quality and similarity to ground truth. The chart highlights distinct structural and perceptual restoration strengths across the tested approaches. Then there's MPRNet [26], which uses a multi-stage CNN, giving a decent overall performance that balances complexity and general performance. But still, MFDNet beats it in every category.

As shown in Figure 7, Darker shades represent better scores, enabling quick visual comparison. The chart highlights MFDNet's consistently strong performance across PSNR and SSIM, with clear distinctions among methods in both synthetic and real-world conditions. Lighter shades are preferable for LPIPS, where lower values indicate better perceptual quality. All in all, these numbers really show how MFDNet shines when it comes to reducing glare, keeping detail, and restoring the overall look of real-world images.

To see how MFDNet stacks up, let's compare it to older flare removal methods. Table IX gives a quick rundown of how traditional models, U-Net-based approaches, Transformer-based techniques, and frequency-decomposed models perform across four key areas: how well they remove flares, how well they keep details intact, their computational demands, and how easily they can be scaled up or adapted.**** 

MFDNet’s Superior Attributes: 

Overall, MFDNet offers a more balanced and efficient approach to flare removal, especially in situations that demand quick and reliable performance.

Future Research Directions 

Although substantial progress has been made in flare removal, current solutions face significant limitations when applied to real-world scenarios. This section highlights ongoing challenges and outlines key directions for future research aimed at enhancing both model robustness and practical deployment.

Flare removal techniques have advanced significantly with deep learning, supported by diverse datasets and improved evaluation methods. Among the compared approaches, MFDNet demonstrates strong performance in flare removal, detail preservation, and computational efficiency. However, challenges such as real-world generalization, extreme flare conditions, and real-time deployment remain, highlighting the need for more robust, adaptable, and efficient models.