GDT-Net: A Model Based on Deep Learning for Water Surface Garbage Detection
Yanran Pang, Bo Qin · 2021
Traditional water surface garbage detection algorithms have the shortcomings of low detection accuracy, slow speed, and susceptibility to interference. Therefore, this paper proposes a new single-stage object detection model called GDT-Net, which is divided into three parts: feature extraction network, feature fusion network, and detection head. In the feature extraction network part, a new cross-stage multi-scale partial feature extraction module (CSMP) is designed by combining CSPNet and Res2Net. It enhances the learning ability of CNN without increasing the number of parameters. The feature fusion network uses PANet and adds a newly designed multi-scale atrous convolutional enhancement module (MACM) and an improved attention model (IAM) to PANet to focus and enhance the valuable features while enhancing the semantic information of shallow features. The experimental results show that the Precision and Recall of GDT-Net are 89.3% and 92%, respectively, and the mAP reaches 91.1%, and the FPS is 38, which is better than the mainstream object detection networks such as Faster RCNN and YOLOv3 in terms of detection accuracy and speed.