Bias and Fairness in Low Resolution(Very) Image Recognition
Kotti, Sasikanth
Bias and Fairness in Low Resolution(Very) Image Recognition by Sasikanth Kotti - IIT Jodhpur Department of Computer Science and Technology 2023 - x,33p. HB
Recent image recognition algorithms showed great performance, which was possible due to advancements in deep learning methods. These methods find application in a variety of scenarios such as recognizing species, surveillance, identifying missing persons, and identifying objects from drones during floods. However, current methods assume the availability of images of very high resolution. Obtaining images of high resolution is not always possible due to the large distance of the object and inherent limitations in the acquisition device, such as the camera. Hence, it is important to develop robust algorithms that can also work with low-resolution and very low-resolution images. It is also essential that these algorithms are not biased and are fair to different sub-groups.
Recent algorithms in the literature showed decent improvement in performance. However, bias and fairness were not taken into consideration in the current algorithms, especially in components that use generative models. Hence, there is enormous scope to further improve the performance of image recognition with low resolution along with developing fair algorithms. In this work, we attempted to cover existing literature. We showed experimentally that this problem needs further research for improved performance. We also showed that existing generative models, specifically GANs, are prone to bias and fairness issues and can cause disparate impact. Lastly, we proposed methods and techniques to debias existing generative models. We hope these techniques can be used to develop fair algorithms for low-resolution image recognition.
Department of Computer Science and Technology
Generative Models
MTech Theses
006.3 / K877B
Bias and Fairness in Low Resolution(Very) Image Recognition by Sasikanth Kotti - IIT Jodhpur Department of Computer Science and Technology 2023 - x,33p. HB
Recent image recognition algorithms showed great performance, which was possible due to advancements in deep learning methods. These methods find application in a variety of scenarios such as recognizing species, surveillance, identifying missing persons, and identifying objects from drones during floods. However, current methods assume the availability of images of very high resolution. Obtaining images of high resolution is not always possible due to the large distance of the object and inherent limitations in the acquisition device, such as the camera. Hence, it is important to develop robust algorithms that can also work with low-resolution and very low-resolution images. It is also essential that these algorithms are not biased and are fair to different sub-groups.
Recent algorithms in the literature showed decent improvement in performance. However, bias and fairness were not taken into consideration in the current algorithms, especially in components that use generative models. Hence, there is enormous scope to further improve the performance of image recognition with low resolution along with developing fair algorithms. In this work, we attempted to cover existing literature. We showed experimentally that this problem needs further research for improved performance. We also showed that existing generative models, specifically GANs, are prone to bias and fairness issues and can cause disparate impact. Lastly, we proposed methods and techniques to debias existing generative models. We hope these techniques can be used to develop fair algorithms for low-resolution image recognition.
Department of Computer Science and Technology
Generative Models
MTech Theses
006.3 / K877B