Research On Key Technologies For Object Detection In Remote Sensing Images

With the rapid development of remote sensing and deep learning technology,intelligent interpretation of remote sensing images has become a research hotspot.In addition,few-shot object detection,which aims to effectively detect objects of novel classes by a small number of labeled examples,has also attracted more and more researchers’ attention.Object detection and semantic segmentation in remote sensing images are key technologies for intelligent interpretation of massive remote sensing data,and they play an important role in many application scenarios.Few-shot object detection can greatly reduce the cost of data labeling,and further enhance the learning ability of the machine.This paper studies object detection and associated semantic segmentation in remote sensing images.The main works are as follows:1.From the perspective of explicit modeling of local context,an object detection algorithm for remote sensing images based on context enhancement is proposed,which obtains more discriminative multioriented features and improves the model’s ability to distinguish heterogeneous objects with similar appearances.The experimental results on the public dataset verify the effectiveness of the proposed algorithm.2.A semantic segmentation network for remote sensing images based on hybrid asymmetric dilation convolution is proposed.The hybrid asymmetric dilation convolution structure is used to replace the dense dilation convolution in the decoder,which effectively fills in the holes in the receptive field,maintains the semantic consistency of adjacent features,and improves the performance of the segmentation network.3.A new few-shot object detection algorithm fused with GIoU loss is proposed to guide the network to generate predicted bounding boxes that are closer to ground truth,enabling the model achieves better detection results under high IoU matching thresholds.