Low Power Approximate Attention guided ROI Module for Detection of UAV Images

In the world of Unmanned Aerial Vehicles (UAV), accurate object detection by maintaining low power and low area is a major challenge. In error resilient applications like object detection, image processing, and machine learning, where accuracy is not that crucial, approximate arithmetic [4] circuits can be used [5], which reduces the power consumption, and the area, thereby increasing speed and performance. The present work proposes an approximate Attention Guided Region Of Interest (AROI) module of Self attention - Global Feature Fused (SAGFF) network. The proposed approximate AGROI is developed by approximate arithmetic blocks. The results show that AROI module has very less mean error distance and normalized mean error distance. The proposed AGROI module consumes less power by virtue of approximate blocks and is suitable for energy efficient UAV object detection.