Image Dehazing for Object Recognition using Faster RCNN
Bhanu Teja Nalla, Teena Sharma, Nishchal Kumar Verma, Soumya Ranjan Sahoo · 2018
Object recognition in hazy conditions is quite difficult due to illumination variance. The challenge arises in finding out features from such images. Herein, we have proposed the method to deal with such images. The input image needs to be dehazed before applying the recognition algorithms. On the other hand, dehazing a non-hazy image makes it dark resulting in loss of features. Hence, a decision is to be made whether or not the image should be dehazed before recognition. Also, for a very dense haze, even dehazing doesn't help in object recognition. In order to tackle this issue, this paper presents a novel method to quantitatively estimate the amount of haze in the image - also termed as haze degree - using dark channel prior of the input images. We compared our values with the existing method using FRIDA dataset. The estimated haze degree is used to decide whether input image need to be dehazed or not. We use DehazeNet and Faster RCNN for dehazing and recognition, respectively. We test our method on real time hazy images to set a threshold on haze degree to classify the image as light, moderate or densely hazed. We used the Static Scenes dataset from Color Hazy Images for Comparison (CHIC) database to obtain the threshold values.