Melhoramento de Imagens de Neblina

Adaptive Parameter Learning through Convolutional Neural Networks for Image Dehazing

The continuous advances in imaging technology enabled several applications in different scenarios, in particular image analysis and manipulation. Computer vision has addressed different problems related to the analysis and interpretation of images, such as classification, detection and tracking of objects.  However, analyzing and interpreting images are extremely challenging tasks, especially in outdoor environments. 

Images captured outdoors can easily be affected by natural environmental effects, such as variation in lighting, shadow and reflection. However, images with fog can represent an even more complex challenge, as they typically have low contrast, visibility and saturation. Therefore, hazy images can depict a great challenge for several applications related to image analysis and interpretation.

The purpose of image dehazing is to enhance the visual quality of degraded images by reducing the impact of weather conditions. This is achieved by reducing the obscuring effects of these elements, thereby improving clarity, contrast and the overall visibility of objects and details within the image. Image dehazing is particularly crucial in several applications, including remote sensing, surveillance and semantic segmentation, where clear and accurate visual information is essential for a more precise analysis and interpretation.

Well-known image processing techniques such as color adjustment, gamma correction and exposure enhancement methods can be used together, as an Image Signal Processing (ISP) pipeline, in order to dehaze images achieving satisfactory results. However, due to the number of parameters these techniques require, the results are usually focused on a particular set of image features and they require manual parameter adjustments.

Considering that Convolutional Neural Network (CNN) can automatically and adaptively learn features from a dataset, it can be deployed to learn the necessary parameters for the ISP image processing techniques mentioned earlier. The idea is that by learning the parameters, the resulting model can produce an image dehazing technique with more precise visual results.

In this paper we propose an image dehazing approach, removing fog effects from images of outdoor environments, using Convolutional Neural Network (CNN). To achieve this, a CNN regression model was employed to enhance image quality by estimating the optimal parameters for the ISP pipeline, aiming at the enhancement process of hazy images. Next, a series of intensity transformation methods are applied in order to reduce the degradation in hazy images. In the experiments we used two well-established hazy image datasets. Additionally, a comparative analysis was performed, comparing our results with other state-of-the-art algorithms in the field. The results validate the effectiveness of our proposed approach for hazy image restoration, highlighting a significant improvement in accuracy and clarity of hazy images.

The main contributions of our work are summarized as follows:

• We propose a CNN-based approach to learn and estimate the best parameters for a novel adaptive ISP pipeline in the process of enhancing the quality of hazy images. The proposed regression process acquires knowledge from the different fog conditions (like saturation, low contrast, lighting, scattering and blurry in haze), enabling the estimation of the optimal parameters for different hazy images. The proposed enhancement approach presents high accuracy even in different haze conditions;

• We propose a challenging dataset composed by over 4700 hazy images, acquired in urban and natural outdoor environments. The proposed dataset comprises hazy images with intense fog and low visibility and scattering. To the best of our knowledge this is the first dataset of hazy images from urban and natural outdoor environments, composed of aerial images.

Publications:

Acknowledgments:

This work was developed with support from the Motorola, through the IMPACT-Lab R&D project, in the Institute of Computing (ICOMP) of the Federal University of Amazonas (UFAM).

Team:

Laura A. Martinho, Graduate Student at Universidade Federal do Amazonas (UFAM)

João M. B. Calvalcanti, Associate Professor at Universidade Federal do Amazonas (UFAM)

José L. S. Pio, Associate Professor at Universidade Federal do Amazonas (UFAM)

Felipe G. Oliveira, Adjunct Professor at Universidade Federal do Amazonas (UFAM)