NAS -PRNet Neural Architecture Search generated Phase Retrieval Net for Off -axis Quantitative Phase Imaging Xin Shu1 Mengxuan Niu1 Yi Zhang1 and Renjie Zhou1

2025-05-02 1 0 601.86KB 9 页 10玖币

侵权投诉

NAS-PRNet: Neural Architecture Search generated Phase Retrieval

Net for Off-axis Quantitative Phase Imaging

Xin Shu1, Mengxuan Niu1, Yi Zhang1, and Renjie Zhou1, *

1Department of Biomedical Engineering, the Chinese University of Hong Kong, Shatin, New Territories,

Hong Kong, China

*Corresponding author: rjzhou@cuhk.edu.hk

Abstract: Single neural networks have achieved simultaneous phase retrieval with aberration

compensation and phase unwrapping in off-axis Quantitative Phase Imaging (QPI). However, when

designing the phase retrieval neural network architecture, the trade-off between computation latency and

accuracy has been largely neglected. Here, we propose Neural Architecture Search (NAS) generated

Phase Retrieval Net (NAS-PRNet), which is an encoder-decoder style neural network, automatically

found from a large neural network architecture search space. The NAS scheme in NAS-PRNet is

modified from SparseMask, in which the learning of skip connections between the encoder and the

decoder is formulated as a differentiable NAS problem, and the gradient decent is applied to efficiently

search the optimal skip connections. Using MobileNet-v2 as the encoder and a synthesized loss that

incorporates phase reconstruction and network sparsity losses, NAS-PRNet has realized fast and accurate

phase retrieval of biological cells. When tested on a cell dataset, NAS-PRNet has achieved a Peak Signal-

to-Noise Ratio (PSNR) of 36.1 dB, outperforming the widely used U-Net and original SparseMask-

generated neural network. Notably, the computation latency of NAS-PRNet is only 31 ms which is 12

times less than U-Net. Moreover, the connectivity scheme in NAS-PRNet, identified from one off-axis

QPI system, can be well fitted to another with different fringe patterns.

1. INTRODUCTION

Quantitative Phase Imaging (QPI) has been widely applied to biomedical imaging [1, 2] and material

metrology [3, 4]. In off-axis interferometry-based QPI [5] (off-axis QPI), the Optical Path Difference

(OPD) or phase distribution of an object is encoded in a fringe pattern or interferogram by interfering the

object field with a tilted reference field. To reconstruct the phase map from a recorded interferogram, the

conventional approach contains three key steps: (i) retrieving the wrapped phase (valued between 0 to 2𝜋)

from the complex object field (e.g., using the Fourier transform method [6]); (ii) unwrapping the phase

(e.g., using the Goldstein algorithm [7]); and (iii) calibrating the phase or compensating the phase

aberration by using an additional interferogram captured in a sample-free region [8]. Among the three

steps, phase unwrapping is the most time-consuming; moreover, one may fail to obtain the calibration

interferogram when imaging a dense sample. To expedite phase retrieval in off-axis QPI, parallel

computation using sophisticated Parallel Graphics Processing Units (GPUs) or Field Programmable Gate

Arrays (FPGAs) has been applied to accelerate phase unwrapping [9, 10], while the calibration

interferogram is still required. In recent years, neural networks have become attractive alternatives for

achieving phase retrieval with aberration compensation with/without phase unwrapping in off-axis QPI,

such as the widely used U-Net models [11, 12] and the Y-Net [13]. Despite realizing significant

simplification of the imaging system and reduced cost, furthering applying those methods for real-time

phase imaging in off-axis QPI is potentially limited by the relatively large computation latency and phase

retrieval accuracy. It is known that the network inference accuracy and efficiency heavily depend on its

architecture. Therefore, a strategy to identify an optimal network architecture is needed to minimize the

computation latency, while keeping the phase retrieval accuracy high.

Neural architecture search (NAS) [14] is a technique to automatically find an optimal network

architecture from a large architecture search space. NAS generated networks have outperformed manually

designed networks in many tasks, including classification [15-17] and end-to-end dense image prediction

(e.g., semantic segmentation and stereo depth reconstruction) [18-21]. SparseMask [22] is an end-to-end

dense image prediction NAS scheme. SparseMask has a search space that covers different skip

connection strategies from the encoder to the decoder, which enables searching for optimal ways to fuse

low-level features rich in spatial details and high-level features containing semantic information. In

SparseMask, differentiable NAS search strategy [23] is used to relax the architecture search space from

discrete to continuous, which can enable an efficient optimization of skip connections based on gradient

decent. Taking both dense image prediction accuracy and connectivity sparsity into account, SparseMask

attains comparable results while runs more than three times faster compared with the widely-used

Pyramid Scene Parsing Network (PSPNet) [24] on semantic segmentation PASCAL VOC 2012 test

dataset [25].

For achieving accurate phase retrieval with high efficiency, we propose NAS generated Phase Retrieval

Net (NAS-PRNet), as illustrated in Fig. 1. Before generating NAS-PRNet, an intermediate super-network

for phase retrieval (denoted as super-PRNet in Fig. 1(b)), containing all the encoder and the decode

connections, is firstly built up. After training the super-PRNet, the encoder and decoder connections are

pruned according to the connection weights in super-PRNet to obtain the architecture of NAS-PRNet.

Then, NAS-PRNet is trained and tested for phase retrieval. Both super-PRNet and NAS-PRNet are

trained and tested with the same biological cell dataset. The search scheme for NAS-PRNet is modified

from SparseMask but differs in two aspects: (1) encoder features are allowed to propagate into all stages

of the decoder to enlarge the search space, while encoder features in SparseMask are only allowed to

propagate into their corresponding lower-level decoder stages; (2) a global sparsity restrict is adopted to

make the total number of connections as small as possible, while SparseMask uses a sparsity restrict for

each decoder stage on a fixed quantity of connections. Moreover, the network structure of super-PRNet or

NAS-PRNet, including output layer (i.e., adopting a regression head), feature fusing style, kernel size in

convolution, feature depth strategy, etc., is customized for the phase retrieval problem.

2. METHODS AND PRINCIPLES

A. Construction of Super-PRNet

Fig. 1. Phase Retrieval Net (NAS-PRNet). (a) NAS-PRNet retrieves a phase map from an interferogram. (b)

Search scheme for NAS-PRNet. Super-PRNet is the super-network covering the whole search space, and

NAS-PRNet is the searched sparsely connected neural network. Blue circles in the schematics represent

stages of the encoder and decoder, the lines with arrow represent connections from the head stage to the tail

stage, the numbers in the decoder circles represent the number of input features.

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

NAS-PRNet:NeuralArchitectureSearchgeneratedPhaseRetrievalNetforOff-axisQuantitativePhaseImagingXinShu1,MengxuanNiu1,YiZhang1,andRenjieZhou1,*1DepartmentofBiomedicalEngineering,theChineseUniversityofHongKong,Shatin,NewTerritories,HongKong,China*Correspondingauthor:rjzhou@cuhk.edu.hkAbstract:Singleneu...

展开>> 收起<<

NAS -PRNet Neural Architecture Search generated Phase Retrieval Net for Off -axis Quantitative Phase Imaging Xin Shu1 Mengxuan Niu1 Yi Zhang1 and Renjie Zhou1.pdf

共9页,预览2页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

NAS -PRNet Neural Architecture Search generated Phase Retrieval Net for Off -axis Quantitative Phase Imaging Xin Shu1 Mengxuan Niu1 Yi Zhang1 and Renjie Zhou1

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: