Improved Real Time Object Detection Method for Remote Sensing Image Based on YOLOv4
Zhuo Chen, Jian Huang, Hao Xu, Xiushan Zhang · 2021 International Conference on Computer Information Science and Artificial Intelligence (CISAI) · 2021
High resolution remote sensing image contains rich information. Object detection technology is the core technique of remote sensing image analysis. Aiming at the problems of slow object detection speed and low performance of dense small object detection in remote sensing image, an improved real-time remote sensing image object detection method based on YOLOv4 is proposed. This method designs an image segmentation algorithm for high-resolution remote sensing image, which reduces a mass of information loss in the process of image scale transformation. The feature of remote sensing image is extracted by CSPdarknet53 network and modified by inserting the SE block in attention mechanism to strengthen the effective channels, and the information is effectively utilized by SPP and PANet modules. By modifying the loss function and adding a hyper-parameter to suppress the influence of imbalance between the object and background class on the training process. Experiments verify the effectiveness of ImYOLOv4, showing that our proposed approach can outperform the baselines.