Computer Engineering



  • Online:2020-03-06 Published:2020-03-06

Abstract: In remote sensing, airplane target detection is frequently faced with problems including complex background, multiple non-target areas, protruding scale differences of planes and a lack of samples.To overcome these obstacles, a detection model is proposed, which includes two kinds of deep neural networks: full convolution neural network (FCN) and convolution neural network (CNN).The processing of the model mainly includes: candidate region selection, CNN feature extraction and classification, detection frame fusion and so on.The candidate regions of each object are obtained by clustering FCN segmentation graph with density clustering, which can obtain adaptive size candidate region ;using VGG-16 net to extract high-level features of candidate regions to obtain high-accuracy classification results, and to obtain detection frame;A new detection frame suppression algorithm is proposed to suppress overlapping and false detection detection frames;In this process, image-level labels are used to replace target-level labels for CNN training, and CNN bottom level feature maps are applied to produce pixel-level labels for improved FCN training. Experimentations have demonstrated excellent detection performance and good generalization ability of our approach: the model achieves 95.78% accuracy, 98.98% recall and 0.9735 F1 score.