Copyright © 2026 Authors retain the copyright of this article. This article 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.

@article{209470,
        author = {Sarthak Agrawal and Aryan pawar and Dr. Chaitali Patil},
        title = {Real-Time Low-Light Image and Video Enhancement System},
        journal = {International Journal of Innovative Research in Technology},
        year = {2026},
        volume = {13},
        number = {5},
        pages = {1791-1808},
        issn = {2349-6002},
        url = {https://ijirt.org/article?manuscript=209470},
        abstract = {Night, indoor and backlight images or videos fall into just the bottom sliver of brightness range and in so doing are degraded to have low contrast and a dim, noisy sensor-level output. Images of these qualities are very hard for people to process and very difficult for the computer vision systems which utilize their features, for such tasks ranging from surveillance, to driver assistance systems to medical imaging. Supervised CNN methods require corresponding low-light and normal-light training images which are time consuming to capture/curate and not always feasible, whereas histogram equalization and Retinex models fail with poorly distributed lighting. Here we demonstrate LumiDCE, an entirely zero-reference system that defines the problem of low-light image restoration as prediction of a high order light-enhancement curve for per-pixel calculation and not directly an output image. Through a seven-layer deep CNN utilizing depth-wise separable convolutions and symmetrical skip connections, we predict and then repeatedly apply the parameters of the curve to generate the enhanced result. Training requires absolutely zero reference imagery whatsoever, rather relying upon a 5 term non-reference loss function which guarantees the relative spatial consistencies of outputs along the x and y directions, maintains reasonable illumination levels within every output pixel image, preserves reasonable relative color tones across an output image and on videos preserves per-pixel consistency between output image frames over time. With merely 10,561 trainable parameters (just under 100KB in disk size) versus the more than 79,416 in the baseline curve estimation method, and an extremely fast PyTorch-independent numpy kernel which reproduces the recursive function precisely. Our implementation of LumiDCE comes packaged as a full pipeline with the full ability to perform an operation: as a Flask based REST service, from within a web-based image editing suite that can display real-time comparison and also histograms on the live outputcanvas, a live sixty frame per second webcam pipeline, or simply one image and one command line to enhance folders orMP4 videos. We see on the standard benchmarks the spectral Shannon entropy increase from 2.31bits to 5.41bits, the average luminance value increase by roughly one order of magnitude from 0.013to 0.318, and a TEN-FOLD improvement in average luminance contrast along with 5 robust testing runs (via automation) of the REST API endpoints, PyTorch networks, and Numpy kernel within 1.68 seconds. In contrast to many enhancement methods,we report and do not intentionally suppress noise amplification which can occur.},
        keywords = {Curve estimation, depthwise-separable convolution, low-light image enhancement, video enhancement, zero-reference learning},
        month = {October},
        }