Mixed Weighted Classification Loss Function and TFcaNet Attention Mechanism for Remote Sensing Images Object Detection

Abstract:
With the development of intelligent technology, the intelligent processing of remote sensing images has become a research hotspot and the object detection for remote sensing images is becoming more and more extensive. The characteristics of sample imbalance in categories, numerous small objects, and blurry object edges that remote sensing images contains bring huge challenges to object detection. Aiming at the characteristics of sample imbalance in categories of remote sensing images, we designed a mixed weighted classification loss function. At the same time, the TFcaNet attention mechanism is proposed to improve the capabilities of small objects detection and localization of the deep learning model. We create a model named YOLOv5-FRD based on YOLOv5. Through training and testing, the mAP on DOTA dataset reached 73.53%. Compared with the current mainstream one-stage models, YOLOv5-FRD achieves the second place in mAP and the highest AP in the most categories.