Investigating sanitized and controlled image dataset to train deep convolutional neural networks for remote object detection on the field

Abstract

Performing specific object detection and recognition at the imaging sensor level, raises many technical and scientific challenges. Today state-of-the-art detection performances are obtained with Deep Convolutional Neural Network (CNN) models. However reaching the expected CNN behavior in terms of sensitivity and specificity require to master the training dataset. We explore in this paper, a new way of acquiring images of military vehicles in sanitized and controlled conditions of the laboratory in order to train a CNN to recognize the same visual signature with real vehicles in realistic outdoor situations. By combining sanitized images, counter-examples and different data augmentation techniques, our investigations aim at reducing the needs of complex outdoor image acquisition. First results demonstrate the feasibility to detect and classify, in real situations, military vehicles by exploiting only signatures from miniature models.

Method for training deep neural networks in vehicle detection using drone-captured data and background synthesis

“…As dataset A1 alone yielded worse results than the reference model R, in this training run a combination of real background images and the transformed ones from dataset A1 has been investigated. This hybrid approachs yields better results, the precision has been improved by 18 6. UC -BA: By adding "battlefield" access (BA), the model precision is maximized, but at the expense of a significant drop in recall compared to the reference model R.…”

Automated method for generating datasets of infrared or visible images for context-specific training of deep neural network-based object detectors

“…The result of the acquisition are 72 key images per vehicle, corresponding to 18 different yaw angular positions at 4 different pitch angles. This acquisition mode mimics the operation of the ISL optical bench, presented in, 3 with the difference that this time a real vehicle is "scanned", parallely in the infrared and visible spectral domain with varying backgrounds. After background removal, the images series is transformed into a sanitized dataset of infrared and visible images to be used for data augmentation and subsequent CNN training.…”

“…The data-augmentation process is the same as already published. 3 Through the use of data augmentation, one can synthetically expand a dataset, producing a myriad of image variations using diverse transformation combinations. This subsection will give an overview of the augmentation workflow, offering insights into the CNN training and evaluation procedures discussed in section 4.…”

“…Recent publications 3,22,23 demonstrated the interest of data augmentation for improving the generalization capability of CNN trained on synthetic images.…”

“…Three categories of Albumentation 15 transforms have been defined for this study, that particularly worked well to detect and identify real vehicle images within our previous study. 3 They are executed in the following order:…”