Light-YOLOv7: Lightweight Ship Object Detection Algorithm Based on CA and EMA

Ship object detection technology is essential for ensuring the safety of maritime traffic and unmanned ships, among other fields. Most ship detection algorithms are deployed in unmanned surface Vehicles, so the detection algorithm should meet the two aspects of lightweight model and high detection precision. However, most of the current algorithms in the pursuit of precision make the model have an excessive number of parameters and the lightweight model cannot meet the detection precision. To address this challenge, this paper improves the current relatively advanced YOLOv7 algorithm and proposes Light-YOLOv7. On the basis of YOLOv7, The GhostNet with CA attention mechanism is employed as the backbone network for feature extraction in the detection algorithm, which reduces network parameters and computation. The E-ELAN module in the neck is combined with the EMA attention mechanism, forming the EMA-E-ELAN module, the EMA-E-ELAN module enhancement network is mainly capable of sensing and multi-scale feature extraction. The auxiliary head in YOLOv7 was abandoned, and only the lead head was used to complete the final loss calculation and positive and negative sample matching. Light-YOLOv7 algorithm reduced the number of parameters by 45% and the computational load by 57.1%, as demonstrated by experimental results. The precision and mAP@0.5 values are almost the same as those of the original algorithm, which are 97.3 % and 98.3 %. With a significant reduction in the number of parameters and calculations, the precision and mAP@0.5 of the Light- YOLOv7 algorithm remain equal to those of the original algorithm. The lightweight of the model is realized.