Image classification is a computer vision task that involves categorizing images into certain predefined classes based on their visual features. In certain real-world scenarios where images are captured in varying conditions, real-time classification needs to be done under artificial or environmental stressors such as occlusion, camouflage, image distortions etc. The state-of-the-art image classification models require large amounts of data to build a robust and resilient classifier unaffected by stressors. We propose a supervised low-shot learning approach to improve image classification on a limited dataset in a stressed environment by incorporating shape-based feature representations along with the high-level CNN image features. In this thesis we show that the proposed fusion model improves upon the benchmark classification accuracy tested on a dataset of military vehicles under varying battlefield stressors. We visually represent the model performance using F1-scores and ROC-AUC plots as performance metrics.