研究目的
To improve the accuracy of multiobject detection in remote sensing images by proposing a scale adaptive proposal network (SAPNet) that addresses the challenges of varying object sizes and distributions.
研究成果
The proposed SAPNet framework significantly improves multiobject detection accuracy in remote sensing images by using multilayer RPNs and feature fusion, achieving a mAP of 89.2% on NWPU VHR-10 and 62.9% on DOTA, outperforming existing methods like RICNN and MS-VANs.
研究不足
The accuracy for storage tanks and ships is not very good due to closely spaced objects and shape variations. The method may not perform optimally on all object categories, and the number of RPN layers is application-dependent, with diminishing returns when using more layers.
1：Experimental Design and Method Selection:
The study is based on the faster R-CNN framework, with modifications including multilayer RPNs for generating multiscale object proposals and a detection subnetwork with fusion features. VGG16 and ResNet-101 models are used as backbones.
2：Sample Selection and Data Sources:
Two datasets are used: NWPU VHR-10 (650 positive images for 10 classes, split into 533 for training and the rest for testing) and DOTA (2806 aerial images for 15 classes).
3：List of Experimental Equipment and Materials:
A server with Intel Xeon CPU E5-1620 v4 at
4：50-GHz and NVIDIA GTX 1080Ti GPU. The Caffe deep learning framework is employed. Experimental Procedures and Operational Workflow:
Training involves image-centric stochastic gradient descent with specific learning rates and iterations (e.g., 50,000 iterations for NWPU VHR-10, 200,000 for DOTA). Proposals are generated using RPNs on conv4_3 and conv5_3 layers, with feature fusion via deconvolution. Non-maximum suppression is applied.
5：Data Analysis Methods:
Performance is evaluated using mean average precision (mAP), precision, and intersection over union (IoU) metrics.
独家科研数据包，助您复现前沿成果，加速创新突破
获取完整内容
