Xi Chen,Zhewen Hou,Christopher A. Metzler,Arian Maleki,Shirin Jalali
Abstract:We investigate both the theoretical and algorithmic aspects of likelihood-based methods for recovering a complex-valued signal from multiple sets of measurements, referred to as looks, affected by speckle (multiplicative) noise. Our theoretical contributions include establishing the first existing theoretical upper bound on the Mean Squared Error (MSE) of the maximum likelihood estimator under the deep image prior hypothesis. Our theoretical results capture the dependence of MSE upon the number of parameters in the deep image prior, the number of looks, the signal dimension, and the number of measurements per look. On the algorithmic side, we introduce the concept of bagged Deep Image Priors (Bagged-DIP) and integrate them with projected gradient descent. Furthermore, we show how employing Newton-Schulz algorithm for calculating matrix inverses within the iterations of PGD reduces the computational complexity of the algorithm. We will show that this method achieves the state-of-the-art performance.
Information Theory,Computer Vision and Pattern Recognition,Machine Learning,Image and Video Processing,Applications
What problem does this paper attempt to address?
The problem that this paper attempts to solve is to restore images in the presence of speckle noise. Specifically, the paper focuses on the speckle noise problem that is prevalent in coherent imaging systems and proposes a method based on Deep Image Prior (DIP) to improve the quality of image restoration. The main contributions of the paper include:
1. **Theoretical contributions**:
- Established the first theoretical upper bound of the Mean Squared Error (MSE) of the maximum - likelihood estimator under the Deep Image Prior assumption.
- This theoretical result reveals the relationship between MSE and the number of parameters in the Deep Image Prior, the number of observations (looks), the signal dimension, and the number of measurements in each group of observations.
2. **Algorithmic contributions**:
- Introduced the concept of "Bagged - Deep Image Prior" (Bagged - DIP) and combined it with Projected Gradient Descent (PGD).
- Proposed a method of using the Newton - Schulz algorithm to calculate the matrix inverse in PGD iterations to reduce the computational complexity.
### Paper background
In coherent imaging systems, speckle noise is a common problem, which can cause a serious degradation in image quality. Speckle noise can be modeled as multiplicative noise, which makes it very challenging to recover the original signal from the observed data. To address this problem, many practical systems adopt multi - view or multi - shot techniques to reduce the impact of noise through multiple observations.
### Model description
The model considered in the paper is as follows:
\[ y = AX_0 w+z \]
where:
- \( y \) is the observation vector,
- \( A \) is the measurement matrix,
- \( X_0 \) is a diagonal matrix, whose diagonal elements represent the complex - valued signal \( x_0 \),
- \( w \) is the speckle noise, assumed to be independent and identically distributed complex Gaussian random variables,
- \( z \) is the additive noise, usually caused by the sensor, assumed to be independent and identically distributed complex Gaussian random variables.
In a multi - view system, the model becomes:
\[ y_l = AX_0 w_l + z_l \]
where \( l = 1,\ldots,L \), and \( L \) represents the number of views.
### Objective
Due to the complex - valued Gaussian characteristics of speckle noise, the phase of the signal cannot be recovered. Therefore, the purpose of the multi - view system is to accurately estimate the magnitude \( |x_0| \) of the signal based on the observed data \( (y_1,\ldots,y_L) \) and the measurement matrix \( A \).
### Method
The paper proposes two main methods to solve the above problems:
1. **Theoretical analysis**:
- By establishing the theoretical upper bound of the MSE of the maximum - likelihood estimator, the influence of different parameters (such as the number of parameters in the Deep Image Prior, the number of views, etc.) on the restoration performance is analyzed.
2. **Algorithmic implementation**:
- Use Projected Gradient Descent (PGD) combined with Deep Image Prior (DIP) for image restoration.
- Introduced the "Bagged - Deep Image Prior" (Bagged - DIP) method, which improves the restoration quality by using multiple DIP models in each iteration.
- Use the Newton - Schulz algorithm to efficiently calculate the matrix inverse in each iteration, thereby significantly reducing the computational complexity.
### Experimental results
The paper verifies the effectiveness of the proposed method through experiments. The experimental results show that, compared with the existing methods, Bagged - DIP combined with PGD and the Newton - Schulz algorithm has a significant improvement in both image restoration quality and computational efficiency.
### Conclusion
Through theoretical analysis and algorithm innovation, this paper effectively solves the image restoration problem in the presence of speckle noise, and provides new ideas and methods for improving the performance of coherent imaging systems.