Vehicle Detection In Remote Sensing Images Using Deep Neural Networks And Multi-Task Learning

Vehicle detection in remote sensing images has attracted remarkable attention for its important role in a variety of applications in traffic,security,and military fields.Motivated by the stunning success of region convolutional neural network(R-CNN)techniques,which have achieved the state-of-the-art performance in object detection task on benchmark datasets,we propose to improve the Faster R-CNN method with better feature extraction,multiscale feature fusion,and homography data augmentation to realize vehicle detection in remote sensing images.For evaluation,we collected several representative datasets for vehicle detection in remote sensing area.And we also conducted extensive experiments on PASCAL VOC datasets for illustrating the generality and robustness of our model.Considering the collection of remote sensing data,we propose homography-based data augmentation to facilitate detection performance.As deep residual networks exhibit strong learning ability,we make use of Res Net50 for feature extraction.Moreover,we constructed multi-scale feature fusion architecture to capture abstract semantic information,which is more friendly to the detection of small objects.Example mining aims to dig out discriminative examples that can better optimize networks.Here we adopted online hard example mining technique to find vehicles that are uneasy to distinguish.For realizing detection in low-resolution images,we propose to utilize Cycle GAN-like architecture to translate LR domain images to HR domain,in an unsupervised manner.In addition,we embed detector as a discriminator in GAN and back-propagate detection loss to generator for better detection performance,which is proven effective in both remote sensing and PASCAL VOC datasets.