Improving YOLOv5n for lightweight ship target detection

A lightweight target detection algorithm based on YOLOv5n is proposed for the current problems of low accuracy rate and high hardware and software limitations of ship target detection for near-shore port and river monitoring. First, the neck network of YOLOv5n uses Ghost convolutional structure instead of ordinary convolution and bottleneck structure in the C3 module to reduce network complexity and computation. Second, the SA attention mechanism is introduced in the backbone network to strengthen the feature extraction capability. Finally, the SIoU loss is used as the localization loss to speed up the model convergence and improve the accuracy of the algorithm. The experimental results show that the improved model volume is reduced by 13.2% and the average accuracy is improved by 3.8% on average. Deploying the algorithm acceleration to the edge device NVIDIA Jetson nano, the inference time is as low as 38ms for a single image with 640×640 resolution, which basically meets the requirements of real-time ship detection.