An Approach For Noise Induced Object Classification Accuracy Improvement

Among various parameters, large scene object detection and classification accuracy depends on image quality. In general, deep neural networks (DNN) are trained to achieve a desired recognition accuracy on a set of targets. However, DNNs become tuned to the training data used and may not generalize to new unseen data artifacts. Classification accuracy of a previously trained DNN is significantly reduced when classification is run on an image altered with additive noise. In this research, we propose image pre-processing to reduce the impact of noise induced low classification accuracy. Our approach consists of applying compressive sensing inspired pre-processing techniques to noisy images. We then compare the object recognition accuracy of a pretrained model on pre-processed noisy images and unprocessed noisy images. We will present our technical method, results, and analysis on relevant synthetic aperture radar data.