A Machine Learning Approach to Unusual Object Detection in Faint Satellite Imagery

The images taken by by satellites contain different artifacts with varying size. These images are used as input for many applications. The analysis of satellite images are based on human interpretation and this creates limitations. Object detection from aerial images are challenging as manual detection of objects are infeasible and time consuming. It requires automatic detection of various features and objects from the satellite images. Automated detection systems are designed to run in computing systems with high computing power. A Convolution Neural Network architecture optimized for embedded systems is presented here. Bottom up strategy is adopted and the design starts from small architecture and gradually build into a larger architecture till the design objectives are met.