GEVE: A generative adversarial network for extremely dark image/video enhancement
Introduction
Definition of dark image/videos is not limited to images taken during night and encomposes all the images captured under different levels of light from various sources and luminance level less than the day time light outside the studio. We define low light videos as indoor photographs without adequate lighting or light which is not visible to eye at night. Images captured under these conditions have lower contrast besides induction of noise in the images due to inadequate and varying illumination. The causes for uneven illuminations may be low level of lighting or weather conditions [1] prevailing, or it can be due to long distance from light source or poor quality of sensors used in the image capturing system. Due to the above said factors the quality of images and the integrity of the images suffer resulting in unnatural look. Aim of dark image enhancement is to bring such images on par with the images taken under extremely dark condition is much difficult because the originally taken images through surveillance cameras are under deteriorated condition of illumination interspaced with uneven lighting. An alternative way to enhance images/videos taken under relative luminance of 0.1 lux or less is proposed. Proposed model for dark image is based on deep learning technology GAN [2].The major issue taken in this research is the Extreme Dark Video Enhancement Network. There are many techniques available in literature but the expected level of enhancement could not be achieved. In order to bring in more enhancement, this paper proposes GAN (“Generative Adversarial Network”) that are based on GEVE model for solving the issues with the image and video enhancements. The remaining section of the paper isorganized as follows. An overview of the existing methods such as contrast enhancements, intelligent enhancements of low lights videos that are available in literature are discussed in Section 2. Section 3 clearly describes the proposed methodology i.e. the GAN and GEVE. The results of the proposed approach and its comparison with existing works are discussed in the Section 4. Section 5 concludes the findings of the research work.
Section snippets
Traditional image contrast enhnacement methods
Traditional techniques used for image enhancement follows the histogram equalization. This method attempts to stretch the intensity evenly over the image to provided on enhanced contrast. Histogram enhancement [3] aims at an adaptive histographic enhancement. Contrast limited adaptive histogram equalization (CLAHE) [4] controls the level of brightness upgradation obtained from HE. HE has limitations as the color contrasts are highly compromised. Color in the darkened areas of the images are not
Methodology
In the Fig. 1, GAN is shown as consisting of two networks namely generating network and discriminating network. The generating continuously generates data which can deceive the discriminating network. At the same time discriminatory network drives continuously to distinguish between actual and fake image/video frame till a balanced state is arrived. In this extremely low light image /Video frame are provided as input to generator. This generator in turn produces an enhanced image. Training of
Limitation and future work
The present GEVE model works well for indoor data with no illumination. Instead of Google Colab, GPU enabled systems can be used to get more running time and batch size, image size can also be increased while training. Our future work concentrates on 3D convolution layer for video enhancement and the unsatisfied case of outdoor videos.
Conclusion
The GAN-based approach that is discussed here is for boosting extremely dim light images/videos. On the other hand GEVE approach is for adaptively enhance image contrast of extremely low light image/video frame. It works well for indoor data without illumination. It is our assessment that GCF and SSIM methods eminently provides a method for boosting dim light images. GLAD-Net, MSRCR and Retinex-Net algorithms provides lower accuarcy of 79%, 84% and 88% respectively whereas the proposed GEVE
Declaration of Competing Interest
None.
References (47)
- et al.
Adaptive histogram equalization and its variations
Comput. Vis. Gr. Image Process.
(1987) - et al.
LLNet: a deep autoencoder approach to natural low-light image enhancement
Pattern Recognit.
(2017) Adaptive histogram equalization and its variations
Comput. Vis. Gr. Image Process.
(1987)- et al.
Getting to know low-light images with the exclusively dark dataset
Computer Vision and Image Understanding
(2019) - et al.
State of the art analysis of low light video noise and various environmental conditions for enhancement
Int. J. Eng. Adv. Technol. (IJEAT)
(2019) - et al.
Generative adversarial nets
- et al.
Contrast limited adaptive histogram equalization image processing to improve the detection of simulated spiculations in dense mammograms
J. Digit. Imaging
(1998) - et al.
Single-scale retinex using digital signal processors
- et al.
Multiscale retinex for color image enhancement
- et al.
An automated multi scale retinex with color restoration for image enhancement