End-to-End United Video Dehazing andDetecting

1 Overview

Taking into account the application of automobile autopilot, author defogging combined with target detection, the establishment of a network end to end.

FIG generating the mist is typically atmospheric scattering model based on:
$I(x)=J(x)t(x)+A(1-t(x))， \tag{1}$ where $I(x)$ represented by FIG fog, $J(x)$ represents the true picture, $A$ and $t(x)$ are two key parameters, respectively represent light transmittance FIG atmosphere, $t(x)=e^{-\beta d(x)}$， $\beta$ atmospheric scattering coefficient, $d(x)$ from the camera to the object is. Therefore, the true picture and can be expressed as $J(x)=\frac{1}{t(x)}I(x)-A\frac{1}{t(x)}+A。 \tag{2}$
On the AOD network draws its application to video defogging, combined Faster-RCNN apply it to the video object detection, as shown shelter of the final model.

2. AOD

AOD network structure as shown below:

AOD formula (1) can be rewritten: $J(x)=K(x)I(x)-K(x)， \tag{3}$其中 $K(x)=\frac{\frac{1}{t(x)}(I(x)-A)+A}{I(x)-1}， \tag{4}$将 $\frac{1}{t(x)}$和 $A$合并到新变量 $K(x)$中。

2.1 Pipline

1. 对输入 $I(x)$提取特征输出 $K(x)$，
2. 应用公式(3)输出清晰图像。

3 AOD应用到视频去雾

Since AOD for defogging a single image, which the authors has been improved to handle video defogging problems, the main problem is that the mixing (temporal fusion) consecutive frames . Because successive frames intrinsically linked, so the use of multi-frame coherence of video defogging have great prospects.

Three Strategies 3.1 Hybrid consecutive frames

The authors also 5 (later explain why 5) pictures input to the network, in three different phases which are fused, analyzed and compared their results

• I-Level Fusion: will cancatenate five branches in the input stage.
• K-Level Fusion: the five branches of the picture feature maps were concatenate the K estimation phase.
• J-Level Fusion: The fused five branches wherein the output stage.
  
  
  
  

The author of AOD as the initialization parameter values ​​to facilitate the training model.

3.2 The choice of hyper-parameters

By contrast the final test, as the number of input select five consecutive frames (3 too, seven consecutive frames not so much caused by the bite Information) in 3,5,7; K-Level Fusion selected as a feature fusion strategy.

4 loss function

Select the MSE as a function of loss

5 Target Detection

slightly