AUTHOR=Zhu Qiang , Ma Ke , Wang Zhong , Shi Peibei 

TITLE=YOLOv7-CSAW for maritime target detection

JOURNAL=Frontiers in Neurorobotics

VOLUME=Volume 17 - 2023

YEAR=2023

URL=https://www.frontiersin.org/journals/neurorobotics/articles/10.3389/fnbot.2023.1210470

DOI=10.3389/fnbot.2023.1210470

ISSN=1662-5218

ABSTRACT=Target detection in maritime search and rescue operations is crucial for ensuring personnel safety and preventing property loss. However, current target detection algorithms often suffer from low detection rates and high false negative rates due to the complex maritime environment and the small size of most targets. These algorithms also lack robustness and generalization. To tackle these issues, we propose an improved maritime search and rescue target detection algorithm based on YOLOv7, named YOLOv7-CSAW.Firstly, we employ the K-means++ algorithm to determine the optimal size for prior anchor boxes, which facilitates an accurate match between anchor boxes and actual objects. Subsequently, we introduce the C2f module, which ensures a lightweight model while obtaining richer gradient flow information. In tandem, we add the non-parameter simple attention module (SimAM) to augment the model's ability to perceive small target features. Additionally, we have enhanced the feature fusion network to an adaptive feature fusion network (ASFF). This network compensates for the lack of high-level semantic features in small targets. Lastly, we utilize the wise intersection over union (WIoU) loss function, which effectively addresses large positioning errors and missed detections, thereby enhancing the model's generalization ability. Our algorithm underwent extensive testing and validation on a maritime search and rescue dataset, using YOLOv7 as the baseline model. Compared to traditional deep learning algorithms, our proposed algorithm exhibits significantly superior detection performance. We observed a mean average precision (mAP) improvement of 10.73% over the baseline model. Consequently, the accuracy and robustness of small target detection in complex scenes are significantly improved.