研究目的
To improve the detection accuracy of small objects in remote sensing images by designing a network with a deconvolution layer after the last convolution layer of the base network.
研究成果
The DeconvR-CNN model significantly improves the detection accuracy of small objects in remote sensing images compared to Faster R-CNN, achieving a higher mean average precision (mAP). The use of a deconvolution layer effectively recovers information lost in pooling layers, making the model particularly useful for small object detection.
研究不足
The method's effectiveness is primarily demonstrated on a specific dataset of remote sensing images containing ships and planes. The generalizability to other types of small objects or different imaging conditions is not explored.
1：Experimental Design and Method Selection:
The DeconvR-CNN model is designed with a deconvolution layer added after the last convolution layer of the base network to recover small object information lost in pooling layers.
2：Sample Selection and Data Sources:
A dataset of 2400 remote sensing images containing ships and planes is used, with 1600 images for training and 800 for testing.
3：List of Experimental Equipment and Materials:
A GTX 1080 GPU is used for inference.
4：Experimental Procedures and Operational Workflow:
The model is trained for 70k iterations with a learning rate of
5：001 for the first 50k iterations and 0001 for the remaining 20k iterations. Data Analysis Methods:
The mean average precision (mAP) is used to evaluate the performance.
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
