Generative Adversarial Networks (GANs) for object detection under rainy conditions

• Main Author: Teo, Oliver Kwok Rong
• Other Authors: Soong Boon Hee
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2020
• Subjects: Engineering::Computer science and engineering::Computing methodologies::Artificial intelligence; Engineering::Electrical and electronic engineering
• Online Access: https://hdl.handle.net/10356/139577
• Institution: Nanyang Technological University

Artificial intelligence is rapidly developing nowadays. Especially, now with the theory of deep learning, object detection is getting more accurate and faster. The complexity of deep learning techniques has dramatically increased the accuracy of object detection. However, during rainy days, object detection methods will drastically drop in accuracy due to the rain droplets on camera lens. Existing methods to de-rain image for object detection such as CycleGAN had been used to remove rain in an image successfully, by translating an image into a rain free image. However, it was unable to remove images that are affected by rain droplets onto the camera lens. To solve this problem, we have developed a neural network for removing rain droplets from single images. We will be using adversarial training to implement an attentive generative network. The main idea is to make the generative and discriminative network more attentive to areas which are affected by rain droplets. The generative network will pay more attention to areas that are affected by rain and surrounding structure by injecting visual attention on the raindrop region. The generative network will firstly try to produce an attention map which will be the most crucial part of the network. This attention map will use recurrent neural network consisting of deep residual networks (ResNet). The discriminative network will be able to assert restored regions local consistency. I will always also be implementing a Faster Region-Convolutional Neural Network (R-CNN) for object detection model to compare the accuracy. A single image is able to be de-raindrop successfully and achieve a more accurate object detection accuracy on de-raindrops image. It was shown that object detection on images that are affected by rain higher false detection as compared to de-raindrops image. What can be done next is to remove the rain streaks of an image or dehazing of an image as many other factors still affect an image during rainy conditions.