Rain Convolutional Dictionary Network (RCDNet)Dengan Frequency Hint Untuk Menghilangkan Noise Hujan Pada Citra Digital

Anggraini, Merisa Anggi (2026) Rain Convolutional Dictionary Network (RCDNet)Dengan Frequency Hint Untuk Menghilangkan Noise Hujan Pada Citra Digital. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5002221152-Undergraduate_Thesis.pdf Restricted to Repository staff only Download (26MB) | Request a copy

Abstract

Citra digital yang diambil pada kondisi hujan sering mengalami penurunan kualitas visual akibat kemunculan rain streaks yang mengaburkan informasi visual dan menurunkan kinerja sistem pengenalan objek. Salah satu metode Single Image Deraining yang banyak digunakan untuk mengatasi permasalahan tersebut adalah Rain Convolutional Dictionary Network (RCDNet). Namun, RCDNet masih memiliki keterbatasan karena hanya memanfaatkan informasi lokal pada domain spasial sehingga kurang efektif dalam membedakan pola hujan dan tekstur asli latar belakang. Penelitian ini modifikasi arsitektur RCDNet pada bagian M-Net dengan mengintegrasi modul FrequencyHint. Modul ini memanfaatkan transformasi Fourier untuk mengekstraksi informasi pada domain frekuensi dan memberikan petunjuk frekuensi global yang membantu proses pemisahan karakteristik hujan dari objek asli. Selain itu, modul FrequencyHint menerapkan mekanisme spectral gate bottleneck untuk menyaring komponen frekuensi secara adaptif dengan tetap menjaga efisiensi model. Hasil eksperimen menunjukkan bahwa model RCDNet yang diintegrasikan dengan modul FrequencyHint berhasil melampaui performa model baseline. Secara spesifik, pada pengujian menggunakan benchmark dataset sekunder dengan model RCDNet yang diintegrasikan dengan FrequencyHint, model ini memberikan peningkatan PSNR sebesar 0,85% dB dan SSIM sebesar 0,08% pada dataset Rain100L (hujan ringan), serta peningkatan PSNR sebesar 1,89% dB dan SSIM sebesar 0,76% pada dataset Rain100H (hujan lebat). Pengujian pada dataset primer sebanyak 487 citra hujan nyata menunjukkan penurunan skor BRISQUE sebesar 13,82% tanpa mengurangi ketajaman detail gambar.
======================================================================================================================================
Digital images captured under rainy conditions often suffer from degraded visual quality due to the presence of rain streaks, which obscure visual information and reduce the performance of object recognition systems. One of the widely adopted methods for addressing this problem is the Rain Convolutional Dictionary Network (RCDNet). However, RCDNet still has limitations because it primarily exploits local information in the spatial domain, making it less effective at distinguishing rain patterns from the original background textures. This study modifies the M-Net architecture of RCDNet by integrating a FrequencyHint module. The proposed module utilizes the Fourier Transform to extract frequency-domain information and provides global frequency cues that facilitate the separation of rain characteristics from the underlying scene content. Furthermore, the FrequencyHint module employs a spectral gate bottleneck mechanism to adaptively filter frequency components while maintaining computational efficiency. Experimental results demonstrate that the RCDNet model integrated with the FrequencyHint module consistently outperforms the baseline model. Specifically, on benchmark datasets, the proposed model achieves PSNR improvements of 0,85% and SSIM improvements of 0,08% on the Rain100L dataset (Light Rain Dataset), as well as PSNR improvements of 1,89% and SSIM improvements of 0,76% on the Rain100H dataset (Heavy Rain Dataset). Moreover, evaluation on a primary dataset consisting of 487 real-world rainy images shows a 13,82% reduction in the BRISQUE score without compromising image detail preservation.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Single Image Deraining, Rain Convolutional Dictionary Network(RCDNet), FrequencyHint, transformasi Fourier, rain streak removal, Fourier Transform
Subjects:	Q Science > QA Mathematics > QA336 Artificial Intelligence Q Science > QA Mathematics > QA404 Fourier series Q Science > QA Mathematics > QA76.6 Computer programming. Q Science > QA Mathematics > QA76.9.I52 Information visualization T Technology > TA Engineering (General). Civil engineering (General) > TA1637 Image processing--Digital techniques. Image analysis--Data processing.
Divisions:	Faculty of Science and Data Analytics (SCIENTICS) > Mathematics > 44201-(S1) Undergraduate Thesis
Depositing User:	Merisa Anggi Anggraini
Date Deposited:	31 Jul 2026 03:45
Last Modified:	31 Jul 2026 03:45
URI:	http://repository.its.ac.id/id/eprint/140629

Actions (login required)

View Item