基于参数量和感受野可调的遥感目标检测算法

In order to solve the problem of multi-scale and poor real-time performance in optical remote sensing image detection, a remote sensing target detection algorithm based on the adjustable parameter number and receptive field is proposed, which can not only reach high detection accuracy, but also achieve real-time performance. Based on the faster region-convolution neural network, a receptive field adjustable module and a channel number adjustable module are designed to improve the accuracy and speed respectively. At the same time, in order to reduce parameter redundancy, the dimension of the full connection layer changes dynamically according to the number of target categories. Experimental results on remote sensing datasets of DIOR, show that the proposed method is higher than all the comparisons with the highest accuracy, and the detection speed is higher than Faster R-CNN. When our algorithm achieved highest speed, it can achieve real-time requirement with proper precision.