Problems such as small object and helicopter (and ship) mutual occlusion are difficult to detect in complex backgrounds, which poses great challenges to the accuracy and real-time detection of helicopter and ship. The application of the YOLOV3 algorithm with high real-time performance to the detection of helicopter and ship under complex backgrounds cannot reach a satisfactory level. This paper has made four improvements on the basis of YOLOV3 algorithm: l)Replace the downsampling in the backbone network with dilated convolution, maintain a higher resolution and a larger receptive field, improve the accuracy of the model for small object detection; 2)Introduce channel attention module to extract more semantic information of target objects; 3)optimizes the non-maximum suppression algorithm by linear declining the confidence score to improve the model’s ability to detect occlusion helicopter(and ship); 4) IOU algorithm is optimized and solved When IOU is used as Loss, Loss =0 due to the disjointness between the prediction box and the groundtruth, which makes it impossible to optimize. The results show that the improved YOLOV3 can detect small object and occlude helicopter(and ship), and improve the detection accuracy from 74.04% to 89.04% under the premise of ensuring real-time performance.
Aerospace target detection based on complex background
28.09.2020
1791737 byte
Aufsatz (Konferenz)
Elektronische Ressource
Englisch
Freight car target detection in a complex background based on convolutional neural networks
SAGE Publications | 2019
|Aerospace 2020. Volume 3 - Background Papers
NTIS | 1997
TIBKAT | Nachgewiesen 3.2004 - 5.2006; damit Ersch. eingest.