【Objective】Aiming at the problem that the general target detector is not directly applied to remote sensing image detection, a remote sensing image target detection algorithm based on improved RetinaNet is proposed. [Method] The algorithm combines the advantages of dynamic selection of down-sampling blocks and convolution kernels. Firstly, the improved downsampling module (IDM) on the base feature extraction network ResNet50 was introduced, which performed multiple down-sampling processing on features. Then the spatial receptive field was dynamically selected by using the convolution kernel selection mechanism to model the multiscale semantic information of the scene. Finally, the classification and regression results of the target object were obtained.【Result】Experimental results show that the proposed method improves mAP by 3.2 percentage points on the large-scale remote sensing image object detection dataset DOTA compared to the orignal RetinaNet network.【Conclusion】The mechanism of introducing a downsampling module and dynamically selecting the size of the convolution kernel has improved the recognition ability of multi-scale remote sensing targets to a certain extent.