A Deep Lightweight Convolutional Neural Network Method for Real-Time Small Object Detection in Optical Remote Sensing Images

The existing object detection algorithms mainly detect large objects. There are few kinds of researches on small objects. And the small object detection accuracy is low, which is difficult to satisfy the real-time requirement. Therefore, this paper proposes a deep lightweight convolutional neural network method for real-time small object detection in optical remote sensing images. Firstly, we build a data set with small objects. The object occupies a very small proportion in the image. There are some interferences such as truncation and occlusion in this data set, which can better evaluate the advantages and disadvantages of the detection method of small objects. Secondly, combining with the region proposal network, we propose a method for generating high-quality candidate boxes of small objects to improve the detection accuracy and speed. Meanwhile, two new learning rate strategies are proposed to improve the performance of the model and further improve the detection accuracy. Experimental results show that the proposed method is an effective small object detection algorithm and achieves real-time detection effect compared with other state-of-the-art object detection methods.