부산대학교 도서관

In recent years, impressive results have been achieved in remote sensing image object detection algorithms based on deep learning. However, multi-spectral remote sensing images have features such as complex backgrounds and small object scales. The traditional object detection methods are difficult to achieve good results, especially for the weak object, which cannot be precisely located and accurately identified. To improve the feature extraction capability and target detection performance, an optimization study is carried out in this paper, and an Efficientnet-YOLOv3i method for target detection and classification of multispectral remote sensing images based on dynamic convolutional neural networks is proposed. For the backbone feature extraction network, which extracts a small amount of effective information and the feature map has a weak ability to characterize the information, A dynamic convolutional neural network module is introduced that increases neither the depth nor the width of the network, and improves the performance of the model by noticing the aggregation of multiple convolutional kernels with convolutional kernels sharing the same kernel size, input and output dimensions. In this study, the effectiveness of this algorithm was tested using RSOD and TGRS-HRRSD remote sensing image datasets. In the EYOLOv3i algorithm proposed in this study, the mean Average Precision (mAP) value is improved by 1.93% and the mean Log Average Miss Rate (mLAMR) value is reduced by 5.00% compared with the EYOLOv3 algorithm in the RSOD remote sensing image dataset. The mAP value increased by 2.90% and the mean log Average Miss Rate (mLAMR) value is reduced by 5.69% compared to the EfficientNet-YOLOv3 algorithm in the TGRS-HRSD remote sensing image dataset.