Proceedings Article | 11 October 2023
KEYWORDS: Detection and tracking algorithms, Object detection, Convolution, Defect detection, Data modeling, Feature extraction, Education and training, Algorithm development, Safety, Performance modeling
To address the issue of imbalanced complexity and precision in series models of YOLO, this paper proposes a lightweight object detection algorithm based on YOLOv5s as the baseline model, namely YOLOv5-SFD. Firstly, the PANet structure of YOLOv5s is simplified to the basic FPN, and a Locally Dense Connected Convolution (LDCC) is proposed to improve the problem of feature information loss during downsampling by cheap feature extraction and feature reuse. Secondly, the C3 module of the original model is improved based on LDCC and combined with the C3Ghost module to reduce the weight of the model. Finally, a bottleneck residual block, BLS (Bottleneck LDCC-SE), based on LDCC and Squeeze-and-Excitation attention is introduced to enhance the feature extraction capability of the network. On the dataset NEU-DET, compared to the baseline model, the precision, recall, F1 score, mAP@.5 and mAP@.5:.95 is 73.4%, 84.7%, 77%, 91.5% and 52.5%, respectively, which represents an improvement of 3.2%, 2.2%, 2%, 11.8%, and 4.5%. In addition, the number of parameters is reduced by approximately 47%, and the model size is only 55% of the baseline model. The experimental results demonstrate that YOLOv5-SFD has lower model complexity and higher detection accuracy, which meets the requirements of lightweight networks.
