《電子技術(shù)應(yīng)用》
您所在的位置:首頁 > 通信與網(wǎng)絡(luò) > 設(shè)計應(yīng)用 > 基于多層次特征提取的輕量級超分辨率重建算法
基于多層次特征提取的輕量級超分辨率重建算法
信息技術(shù)與網(wǎng)絡(luò)安全 5期
竺可沁1,,2,,林珊玲2,,3,林志賢1,,2,,3,,郭太良1,,2
(1.福州大學(xué) 物理與信息工程學(xué)院,,福建 福州350116; 2.中國福建光電信息科學(xué)與技術(shù)創(chuàng)新實驗室,,福建 福州350116,;3.福州大學(xué) 先進(jìn)制造學(xué)院,福建 泉州362200)
摘要: 針對目前基于深度學(xué)習(xí)的超分辨率算法特征提取較為單一,、結(jié)構(gòu)復(fù)雜且參數(shù)龐大的問題,,提出了一種基于多層次特征提取的輕量級超分辨率重建算法。該算法采用了多層次特征提取的方式,,首先提取圖像的淺層特征,;其次,,使用包含多個并行卷積的深層特征提取模塊提取圖像的深層特征。設(shè)計了一種帶學(xué)習(xí)權(quán)重的多尺度特征融合重建模塊,,以充分利用提取出的多層次信息重建圖像。實驗結(jié)果表明,,其重建圖像的峰值信噪比和結(jié)構(gòu)相似性在多數(shù)情況下領(lǐng)先于目前主流算法,;與對比算法相比,在參數(shù)量和運算時間上均保持領(lǐng)先,,證明了網(wǎng)絡(luò)的輕量化特性,。
中圖分類號: TP391.7
文獻(xiàn)標(biāo)識碼: A
DOI: 10.19358/j.issn.2096-5133.2022.05.006
引用格式: 竺可沁,林珊玲,,林志賢,,等. 基于多層次特征提取的輕量級超分辨率重建算法[J].信息技術(shù)與網(wǎng)絡(luò)安全,2022,,41(5):38-44.
A lightweight super-resolution algorithm based on multi-level feature extraction
Zhu Keqin1,,2,Lin Shanling2,,3,,Lin Zhixian1,2,,3,,Guo Tailiang1,2
(1.College of Physics and Information Engineering,,F(xiàn)uzhou University,,F(xiàn)uzhou 350116,China,; 2.Fujian Science & Technology Innovation Laboratory for Optoelectronic Information of China,,F(xiàn)uzhou 350116,China,; 3.School of Advanced Manufacturing,,F(xiàn)uzhou University,Quanzhou 362200,,China)
Abstract: In order to solve the problem that today′s algorithms based on deep learning have simple feature extraction, complex structure and huge parameters, a lightweight super-resolution algorithm based on multi-layer feature extraction is proposed. The algorithm adopts the multi-level feature extraction method. Firstly, the low feature of the image is extracted. Secondly, the deep feature extraction module containing multiple parallel convolutions is used to extract the deep feature of the image. A multi-scale feature fusion and reconstruction module with learning weights is designed to make full use of the extracted multi-level information to reconstruct images. Experimental results show that the peak signal noise ratio and structural similarity of reconstructed images are better than the current algorithms in most cases. Compared with the comparison algorithm, the number of parameters and operation time remain ahead, which proves the lightweight feature of the network.
Key words : super-resolution reconstruction,;multi-level feature extraction;multi-scale feature fusion,;convolutional neural network,;image enhancement

0 引言

隨著信息技術(shù)的飛速發(fā)展與應(yīng)用,圖像處理技術(shù)已經(jīng)成為信息時代的關(guān)鍵核心技術(shù)之一,。數(shù)字圖像在醫(yī)學(xué),、監(jiān)控,、遙感等領(lǐng)域得到了廣泛的應(yīng)用,人們對圖像質(zhì)量的要求也越來越高,。但是由于照片成像質(zhì)量以及保存條件的限制,,使得圖像往往會丟失很多細(xì)節(jié)且分辨率較低,不利于后續(xù)對圖像信息的進(jìn)一步處理,。圖像超分辨率重建(Super Resolution,,SR)技術(shù)可以將一幅低分辨率(Low Resolution,LR)圖像重建為高分辨率(High Resolution,,HR)圖像,。由于其在成本、便利性等方面的顯著優(yōu)勢,,已經(jīng)成為了數(shù)字圖像處理技術(shù)的主要研究內(nèi)容之一[1-2],。





本文詳細(xì)內(nèi)容請下載:http://forexkbc.com/resource/share/2000004245





作者信息:

竺可沁1,2,,林珊玲2,,3,林志賢1,,2,,3,郭太良1,,2

(1.福州大學(xué) 物理與信息工程學(xué)院,,福建 福州350116;

2.中國福建光電信息科學(xué)與技術(shù)創(chuàng)新實驗室,,福建 福州350116,;3.福州大學(xué) 先進(jìn)制造學(xué)院,福建 泉州362200)


此內(nèi)容為AET網(wǎng)站原創(chuàng),,未經(jīng)授權(quán)禁止轉(zhuǎn)載,。