Design and Development of Visibility Restoration Techniques for Weather Degraded Images

Abstract

The visibility of outdoor images is greatly degraded due to the presence of fog, haze, smog, etc. The poor visibility may cause the failure of computer vision applications such as intelligent transportation systems, surveillance systems, and object tracking. To resolve this problem, many image restoration techniques have been developed. These techniques play an important role in improving the performance of various computer vision applications. Due to this, the researchers are attracted toward the visibility restoration techniques. It has been found that the majority of existing techniques suffer from various issues such as edge distortion, color distortion, texture distortion, halo artefacts, gradient reversal artefacts, and poor computational speed. To overcome these issues, various visibility restoration techniques are proposed in this research work. A Dark channel prior (DCP) based visibility restoration technique is implemented by designing a Gain intervention based trilateral filter (GITF) for fog affected images. GITF is able to remove the fog from weather degraded images in an effective manner. It is tested on ten (five benchmarks and five real-life) roadside foggy images. The experimental results reveal that GITF has lesser number of artefacts and preserve more significant edges as compared to the existing restoration techniques. GITF is computationally faster than the existing techniques. Therefore, GITF is more suitable for real-time intelligent transportation systems. Although, GITF outperforms the existing techniques in case of foggy images, it is not so effective against remote sensing hazy images. Therefore, a fourth-order partial differential equation based trilateral filter (FPDETF) based restoration technique is proposed to restore hazy remote sensing images. FPDETF is able to reduce halo and gradient reversal artefacts. It also preserves the radiometric information of restored images. The visibility restoration phase is also refined to reduce the color distortion of restored images. FPDETF is evaluated on ten well-known remote sensing images and also compared with seven well-known existing restoration techniques. Although, FPDETF performs significantly better than the existing visibility restoration techniques. However, it suffers from sky-regions and color distortion, especially in the case of images effected from large weather gradients. Therefore, an Integrated visibility restoration model (IVRM) is proposed to solve the above-mentioned problems. It utilizes DCP, bright channel prior (BCP), and gain intervention filter. BCP is used to solve the sky-region problem associated with DCP based restoration. The gain intervention filter is also used to improve computational speed and edge preservation. IVRM is tested on ten well-known remote sensing images. The simulation results show that IVRM is able to remove halo and gradient reversal artefacts. The designed restoration techniques (i.e., GITF, FPDETF, and IVRM), suffer from noise when transmission map approaches toward zero. Thus, the evaluated atmospheric veil is more than the actual value, (i.e., transmission evaluated by utilizing DCP is lesser as compared to an actual one). As a consequence, the restored color could deviate from the actual object and the restored restored image looks like an artificial image. To overcome this issue, a Modified restoration model (MRM) based DCP is designed and implemented. To further improve the atmospheric veil, a modified joint trilateral filter is also implemented to redefine the transmission map to reduce the color distortion problem. The results reveal that MRM performs effectively across a wide range of weather degradation levels without causing any visible artefacts. The techniques designed so far such as GITF, FPDETF, IVRM, and MRM are not soeffective to preserve the texture details, especially in case of a complex background and large weather gradient image. Therefore, the exploration of new alternatives for designing an effective prior is desirable. Thus, in this research work, two novel channel priors are proposed to evaluate the depth map from weather degraded images. The main advantages of these channel priors over the existing prior are (a) eliminate sky region problem and (b) preserves better texture information of the restored image. These channels are Gradient profile prior (GPP) and Oblique gradient profile prior (OGPP). GPP is designed to remove the haze from remote sensing images. The coarse estimated atmospheric veil is also refined by using guided L0 minimization based filter. Moreover, the visibility restoration is also modified to overcome the over saturation and color distortion problems. Extensive experiments demonstrate that GPP can naturally restore the weather degraded image especially at the edges of sudden changes in the obtained depth map. It can achieve a good effect for single image visibility restoration. GPP is able to evaluate horizontal and vertical edges in a local patch. However, it has been found that many oblique edges are present in an input image. Thus, the standard gradient filter is unable to evaluate the oblique edges. Therefore, in this research work, an Oblique gradient profile prior (OGPP) is designed and developed to efficiently estimate the transmission map and atmospheric veil. The transmission map is also refined by developing a local activity-tuned anisotropic diffusion based filter. Thereafter, image restoration is performed using the estimated transmission function. OGPP has an ability to remove fog from still images in an effective manner. The performance of OGPP is compared with recently developed seven visibility restoration techniques over synthetic and real-life foggy images. The experimental results depict the supremacy of OGPP in removing the fog from still images when compared with the existing techniques. Experimental results reveal that the restored image has little or no artefacts. Thorough extensive analyses, it has been observed that the proposed techniques can effectively suppress visual artefacts for weather degraded images and yield high-quality results as compared to the competitive visibility restoration techniques both quantitatively and qualitatively. Moreover, the relatively high computational speed of the proposed techniques will facilitate these in real-time applications.