Small Target Foreign Object Detection Based on Improved YOLO Network

Abstract

We propose an improved YOLOv4 small target detection method for small target foreign object (FOD) recognition that requires as few detection accuracy, speed, and model volume parameters as possible so that it can be easily ported to other mobile devices such as video surveillance or UAVs in the future. Firstly, the backbone feature extraction network of YOLOv4 is replaced by the MobileNetV2 model, which aims to reduce the number of model parameters. Secondly, the three enhanced feature extraction layers of the original YOLOv4 are increased to four, and a new shallow-scale enhanced feature extraction layer is added, which enhances the characterization capability of the model without increasing the model complexity, making the improved YOLOv3 network structure better for small target detection.