Algorithm on remote sensing target detection based on improved YOLOv3

DOI:	10.11809/bqzbgcxb2023.11.035
Keywords:	deep learning; YOLOv3 algorithm; feature scale; convolutional block attention module; lightweight network; K means+〖KG-*2〗+
Abstract:	Aiming at the low accuracy of YOLOv3 algorithm for small targets such as airplanes and ships in remote sensing image detection, this paper proposes an improved YOLOv3 algorithm for remote sensing target detection. This paper uses the H Swish activation function to replace the ReLU6 activation function in the first layer of the Bottleneck Residual block in MobileNetV2.At the same time, in order to make the network pay more attention to the details of remote sensing images, add a spatial channel attention mechanism in MobileNetV2 can pay more attention to the hidden information of the feature map; The improved MobileNetV2 is used to replace the original Darknet 53 backbone network in YOLOv3. The adjusted Focal loss function is used to replace the original loss function in YOLOv3, and the feature scale is reduced on the basis of the original algorithm, which makes the algorithm processing time shorter. The K means++ clustering algorithm is used to cluster the data set to obtain a set of priori boxes. The experimental results on the DOTA dataset show that the improved algorithm reduces the weight model from 18.8 MB to 8.0 MB, reduces the average detection time from 36.6 ms to 28.42 ms, and increases the mAP_0.5 from 61.7 to 91.3. The algorithm improves the detection speed and accuracy.
Issue:	Vol. 44 No. 11 (2023)
Published:	2023-11-28
PDF HTML