Peningkatan kualitas Gambar Menggunakan Arsitektur U-Net Dengan Mekanisme Deep Supervision Pada Gambar Dengan Tingkat pencahayaan Rendah

Okananta, Riko (2026) Peningkatan kualitas Gambar Menggunakan Arsitektur U-Net Dengan Mekanisme Deep Supervision Pada Gambar Dengan Tingkat pencahayaan Rendah. Undergraduate thesis, UPN Veteran Jawa Timur.

	Text (Cover) Cover.pdf Download (650kB)
	Text (Bab 1) Bab 1.pdf Download (281kB)
	Text (Bab 2) Bab 2.pdf Restricted to Repository staff only until 23 July 2029. Download (1MB)
	Text (Bab 3) Bab 3.pdf Restricted to Repository staff only until 23 July 2029. Download (1MB)
	Text (Bab 4) Bab 4.pdf Restricted to Repository staff only until 23 July 2029. Download (2MB)
	Text (Bab 5) Bab 5.pdf Download (244kB)
	Text (Daftar Pustaka) Daftar Pustaka.pdf Download (184kB)
	Text (Lampiran) Lampiran.pdf Restricted to Repository staff only Download (620kB)

Abstract

Research on low-light image enhancement has become increasingly relevant, as evidenced by the exponential growth of related scientific publications, the emergence of international competitions such as NTIRE since its inception in 2017, and the development of dedicated benchmark datasets such as LOL (Low-Light Dataset). In Indonesia, crimes with the highest frequency include aggravated theft and ordinary theft, which predominantly occur from midnight to early morning when illumination conditions are poor. This limitation reduces the effectiveness of conventional surveillance cameras in identifying digital evidence, making low-light image enhancement a potential solution. However, traditional approaches such as Histogram Equalization remain suboptimal, while existing deep learning models often suffer from instability and excessive model complexity. This study proposes a flexible U-Net-based architecture with contraction (encoder) and expansion (decoder) paths for efficient feature extraction and image reconstruction in lowlight image enhancement tasks. To mitigate information loss caused by downsampling, a multilevel input strategy with pooling combinations is applied at each encoder block to preserve richer feature representations. A Contextual Attention Module is incorporated to enhance global spatial and channel-wise contextual modeling on the fused feature maps from skip connections and the decoder path. Furthermore, a Residual Dense Block Network (RDB-Net), which integrates multiscale components, residual connections, and dense connections, is employed to strengthen feature extraction. Deep supervision is implemented through a modified multilevel fusion output to guide the reconstruction process more effectively. The loss function is further refined by combining Mean Absolute Error (MAE) for pixel intensity accuracy, color loss for inter-channel color distribution consistency, and contrast and structural losses to improve sharpness and structural preservation. Experimental results demonstrate that the proposed U-Net model outperforms several baseline methods in terms of balancing stability, convergence, and image quality, achieving optimal evaluation scores of SSIM = 0.898, PSNR = 24.05, and LPIPS = 0.135 on the LOL Benchmark Dataset.

Item Type:

Thesis (Undergraduate)

Contributors:

Contribution	Contributors	NIDN/NIDK	Email
Thesis advisor	Masdiyasa, I Gede Susrama	NIDN0019067008	igsusrama.if@upnjatim.ac.id
Thesis advisor	Muhaimin, Amri	NIDN0023079502	amri.muhaimin.stat@upnjatim.ac.id

Subjects:

T Technology > T Technology (General) > T385 Computer Graphics

Divisions:

Faculty of Computer Science > Departemen of Data Science

Depositing User:

Riko Okananta Okananta

Date Deposited:

24 Jul 2026 08:09

Last Modified:

24 Jul 2026 08:39

URI:

https://repository.upnjatim.ac.id/id/eprint/57944