CNN FOR LICENSE PLATE MOTION DEBLURRING--阅读笔记

1、We focus on blurred images from a real-life traffic surveillance system, on which we, for the first time, demonstrate that neural networks trained on artificial data provide superior reconstruction quality on real images compared to traditional blind deconvolution methods.

本文专注于现实交通监管系统中的模糊图像，第一次证明了神经网络效果浩宇传统的盲解卷积方法。

2、The training data is easy to obtain by blurring sharp photos from a target system with a very rough approximation of the expected blur kernels, thereby allowing custom CNNs to be trained for a specific application (image content and blur range).

训练数据通过一个对期望模糊核的进行粗略接近的系统产生，产生的这些训练数据用于cnns训练不同应用。

3、Traditional image deconvolution methods are based on a simplified mathematical model of the imaging process which they try to invert, usually in an iterative optimization process.

传统的图像解卷积基于一种简单的数学模型，通过迭代优化将模糊的过程反转。

4、用卷积方法进行图像模糊很简单，但是用于去模糊比较困难，即使知道了模糊卷积时的卷积核也一样。

5、The task of blind deconvolution is to find a latent sharp image x by reversing an imaging process which can be in its simplified form defined as

where g is a blur kernel representing degradation due to motion and lens imperfections, and n is additive random noise.

上面的式子是个模糊卷积的过程，而盲解卷积就是上述过程的反过程，其中x是清晰图像，g是模糊核，n是额外的噪声。

6、Modern blind deconvolution methods select a suitable data prior, which is transformed into a simple regularizer in an optimization problem which is solved by alternately estimating the blur kernel g^ and the latent image x^ [5]. 

当前的盲解卷方法就是选择一个合适的数据先验，将其转换成优化问题中的规则化矩阵，通过交替估计模糊核与潜在的图像解决这个优化问题。

7、We chose a 15 layer network CNN-L15 which was the largest and most successful network for text deblurring reported by Hradisˇ et al.

本文的网络为15层，足够用于文本去模糊了。

8、The network consists only of convolutional and ReLU layers and it contains 2.2 M unique weight parameters, i.e. it occupies approximately 9 MB in memory. See Table 1 for the exact network architecture.

本文这个网络只有卷积层和ReLU层，包含2.2M参数，网络结构如下：

9、The motion deblurring CNN represented as F consists of an input data layer F0, convolutional layers F‘ with their weights represented as convoutional kernels W‘ and their
biases b‘:

运动去模糊CNN的数学模型

10、CNN-L15 was trained on a dataset of fixed-size patches D = fY; Xg, where xi 2 X were sharp patches from a surveillance system and yi 2 Y were respective artificially blurred patches. The size of the blurred input patches was 66 × 66 which corresponds to size of output sharp patches 16 × 16 (see Figure 2).

Fig. 2: An example of training data pair {y i , x i }. Left: network input y i – a random blurred crop. Right: corresponding sharp ground truth x i .

给出了CNN-L15网络的输入和输出的大小，输入为66x66的模糊patch而输出为16x16的清晰patch。

11、All images were obtained from two cameras of a traffic surveillance system (examples are shown in Figure 1) from a road with speed limit of 90 kmph. The first camera captured cars approaching approximately from direction of 37 to 57 degrees while images from the second one were of cars approaching from about 59 to 79 degrees.

所有的图片从使用两个摄像机的交管系统获得，监管路段限速90 kmph，第一个摄像头摄像角度为37-57之间，第二个大概59-79之间。

12、