Please use this identifier to cite or link to this item:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/92326
Title: | 方參數:影像訊號處理器參數推薦基於風格生成對抗網路與U網路 FangParameter: Image Signal Processor Parameter Recommendation Based on StyleGAN2 and UNet |
Authors: | 方郁婷 Yu-Ting Fang |
Advisor: | 傅楸善 Chiou-Shann Fuh |
Keyword: | ISP,參數推薦,深度學習,U-Net,StyleGAN2, ISP,Parameter Recommendation,Deep Learning,U-Net,StyleGAN2, |
Publication Year : | 2024 |
Degree: | 碩士 |
Abstract: | ISP (Image Signal Processor)中文為影像訊號處理器,當光進入手機鏡頭被感應器轉成訊號後,需要經由一系列的影像改進算法(例如白平衡、去噪和去馬賽克以及其他影像增強算法)來將訊號轉變成滿足特定應用或需求的影像,由於影像轉換的複雜性,ISP通常具有許多參數需要調整。
ISP 調參是智慧型手機設計與開發中的一個重要階段,因為手機的硬體特性不同,接收到的光訊號也不同,因此在出廠前都需要調整ISP內的參數,目的是為了讓影像達到最高的品質,這些參數通常由具有經驗的專家手動調整,需要不停地反覆測試來找到最佳的參數,因此調整時間可達數周甚至數個月。 本篇論文提出的方法是以StyleGAN2模型為基礎,結合U-Net的架構,藉由深度學習來達到自動化推薦ISP參數,減少ISP調參所需的人力,並且將調整參數的時間減少到幾個小時以內。 ISP (Image Signal Processor) is a processor that processes the signal from light into images. When light enters the mobile phone camera and is converted into a signal by the sensor, it needs to undergo a series of image algorithms (such as white balance, noise reduction, mosaic removal, and other image enhancement algorithms) to transform the signal into an image that meets specific applications or requirements. Due to the complexity of image transformation, ISP typically has many parameters that require adjustment. ISP tuning is a crucial stage in the design and development of smartphones. Due to variations in hardware characteristics among smartphones, the received optical signals differ. Therefore, adjustments to the parameters within the ISP (Image Signal Processor) are necessary prior to manufacturing, aiming to achieve the highest image quality possible. Typically, these parameters are manually adjusted by experienced experts, involving a process of continuous testing to find the optimal settings. Consequently, the tuning process can take several weeks or even months. Our method is based on the StyleGAN2 (style Generative Adversarial Network 2) model and incorporates the U-Net architecture. Through deep learning, we aim to automate the recommendation of ISP parameters, reducing the human effort required for tuning. This approach also significantly decreases the time needed to adjust parameters to within a few hours. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/92326 |
DOI: | 10.6342/NTU202400230 |
Fulltext Rights: | 未授權 |
Appears in Collections: | 資訊工程學系 |
Files in This Item:
File | Size | Format | |
---|---|---|---|
ntu-112-1.pdf Restricted Access | 6.11 MB | Adobe PDF |
Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.