DHCNN for Visibility Estimation in Foggy Weather Conditions
Akmaljon Palvanov, Young Im Cho · 2018
This paper proposes a new method to estimate visibility range in strong foggy weather conditions on a basis of the Deep Hybrid Convolutional Neural Network (DHCNN). Our method is designed to estimate visibility distance from a digital camera in real-time but by way of using deep networks, it becomes a more challenging task to achieve outcomes quickly. In addition to this, prior to making any prediction, the model needs to pre-process each input so it will produce the desired results. As a consequence, our implemented prototype consists of two main stage: pre-processing inputs and classifier. Each of those stages concatenated sequentially. From the outer perspective, this demonstrates our model's architecture very deep and computationally costly. However, these two stages make our model more robust and help to learn only useful features from inputs. Since the first pre-processing stage identifies Region of Interest (ROI) and removes redundant parts from a high-resolution image and sends forward to classifier just ROI part in lower resolution. We witnessed great accuracy in estimating visibility on not only heavy foggy images but also the classification of hazy images fulfilled very accurately.