An Optimal Retinanet Model For Automatic Satellite Image Based Missile Site Detection

Abstract

Satellite image processing is a manually tedious job and offers scope for automation as part of the information extraction process from satellite images. The process of information extraction involves object detection and one of the challenges is ascertaining the minimum number of images required to train the deep learning model to achieve a certain minimum accuracy. To the best of the authors’ knowledge, work in missile site detection is relatively limited, with an existing exploration of the latest one-shot detection methods, such as RetinaNet, being absent. This work proposes an optimal deep learning model based on the RetinaNet framework and training on a minimal dataset. A comparative analysis with previous work paves the road for future research in one-shot methods and optimally trained models. As part of the study, the key findings are that an optimal training scheme based on a minimal training dataset is possible. This step enables a reduction in training time for the development of an optimal missile site detection model is concerned. One of the many techniques to determine the minimal number of training images required to train the object detection model is plotting the number of training images versus the mean average precision. The same is validated in our work. Further, a hybrid scheme based on the two-model concept is tested wherein one model prioritizes Recall while the other prioritizes Precision. Thus a combination of both models to detect a set of targets provides an optimal framework for object detection. Lastly, the study finds that the single-stage RetinaNet algorithm offers the advantage of balancing speed and accuracy over erstwhile two-stage and other single-stage methods.