Depth and IMU aided image deblurring based on deep learning
Alhawwary, Ahmed (2020-06-23)
Alhawwary, Ahmed
A. Alhawwary
23.06.2020
© 2020 Ahmed Alhawwary. Tämä Kohde on tekijänoikeuden ja/tai lähioikeuksien suojaama. Voit käyttää Kohdetta käyttöösi sovellettavan tekijänoikeutta ja lähioikeuksia koskevan lainsäädännön sallimilla tavoilla. Muunlaista käyttöä varten tarvitset oikeudenhaltijoiden luvan.
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:oulu-202006242663
https://urn.fi/URN:NBN:fi:oulu-202006242663
Tiivistelmä
With the wide usage and spread of camera phones, it becomes necessary to tackle the problem of the image blur. Embedding a camera in those small devices implies obviously small sensor size compared to sensors in professional cameras such as full-frame Digital Single-Lens Reflex (DSLR) cameras. As a result, this can dramatically affect the collected amount of photons on the image sensor. To overcome this, a long exposure time is needed, but with slight motions that often happen in handheld devices, experiencing image blur is inevitable. Our interest in this thesis is the motion blur that can be caused by the camera motion, scene (objects in the scene) motion, or generally the relative motion between the camera and scene. We use deep neural network (DNN) models in contrary to conventional (non DNN-based) methods which are computationally expensive and time-consuming. The process of deblurring an image is guided by utilizing the scene depth and camera’s inertial measurement unit (IMU) records. One of the challenges of adopting DNN solutions is that a relatively huge amount of data is needed to train the neural network. Moreover, several hyperparameters need to be tuned including the network architecture itself.
To train our network, a novel and promising method of synthesizing spatially-variant motion blur is proposed that considers the depth variations in the scene, which showed improvement of results against other methods. In addition to the synthetic dataset generation algorithm, a real blurry and sharp dataset collection setup is designed. This setup can provide thousands of real blurry and sharp images which can be of paramount benefit in DNN training or fine-tuning.
To train our network, a novel and promising method of synthesizing spatially-variant motion blur is proposed that considers the depth variations in the scene, which showed improvement of results against other methods. In addition to the synthetic dataset generation algorithm, a real blurry and sharp dataset collection setup is designed. This setup can provide thousands of real blurry and sharp images which can be of paramount benefit in DNN training or fine-tuning.
Kokoelmat
- Avoin saatavuus [31657]