Low Altitude, Slow Speed and Small Size Object Detection Improvement in Noise Conditions Based on Mixed Training
Jingda Que, Honghong Peng, J. Y. Xiong · Journal of Physics Conference Series · 2019
The low altitude, slow speed and small size object which we call LSS-object for short, such as small UAV (unmanned aerial vehicles) have become a hot issue of air defense security, which is difficult to detect and identify accurately from the image. In this paper, aiming at the problem of LSS-object detection under noise environment, the detection method based on deep learning is proposed. Firstly, a standard training dataset consisting 5 classes of typical objects is constructed. Then, the standard dataset is augmented with noise of different intensity. Finally, YOLO v3 algorithm is used to form a LSS-object detection system which can adapt to environment noise. The training and detection experiments were carried out on the GPU server. After only using the noise-free dataset for training, the mAP(mean Average Precision) of the noise-free test set detection reached 81.07%, but the mAP decreased to 20.68% when the noise variance was 0.03.After adopting the mixed training strategy of the dataset with noise variance of 0.01 and noise-free data, the mAP for the test set detection with noise variance of 0.03 was increased to 70.61%, and the mAP still reached 79.85% in noise-free test set detection. The experiment results show that the mixed training strategy can greatly improve the accuracy in the noisy images detection while maintaining a higher accuracy in noise-free images.