Research On Key Technologies Of Boat Object Services

In recent years,the president of China explicitly proposed the Maritime Power Strategy and the thought of "Caring for the ocean,understanding the ocean,and managing the ocean".In the era of big data,the National Big Data Strategy is also an important component of China’s National Strategies.Marine big data is an important way of "understanding the ocean".However,big data technologies have not been effectively integrated with maritime technologies at present.Current mainstream boat object detection algorithms generally follow the general object detection algorithm,and use traditional CNN network as backbone for feature extraction.This kind of backbone has limited feature extraction capabilities,which makes it difficult to fully utilize the massive information in boat image,resulting in low detection accuracy in complex environments and small object detection;Current mainstream boat object services can only provide offline services through relevant libraries,they cannot provide online services,which makes it difficult to share maritime models and resources.In response to these issues,this paper has carried out the following innovative solutions:(1)The Boat R-CNN boat object detection algorithm was proposed,based on the original Cascade R-CNN detector,using the Swin Transformer backbone with self-attention mechanism and stronger feature extraction capabilities,which is transferred from the NLP field,to replace the traditional ResNet backbone.In combination with boat image features,multiple improvements have been made to anchor boxes,loss functions,and other algorithmic details.Experimental results show that the improved Boat R-CNN algorithm has increased accuracy by 21.8% compared to the original Cascade R-CNN algorithm and 30.4% compared to the mainstream Faster R-CNN algorithm,with no perceivable speed loss for users;(2)The Boat R-CNN+ multi-scale boat object detection algorithm was proposed,further improved from the Boat R-CNN,borrowing the idea of "two-stage algorithms" to improve the original FPN feature pyramid network into two stages,and using various strategies such as copypasting and mosaic enhancement for data augmentation on small objects.Experimental results show that the Boat R-CNN+ algorithm has further increased the small object detection accuracy by 18.3% compared to the Boat R-CNN algorithm,and has increased the small object detection accuracy by 53.5% compared to the mainstream Faster R-CNN algorithm;(3)A resource sharing service platform for boat models was constructed.The platform provides normalized services through a universal API interface,supports multi-user access,and external applications can call shared resources in the platform through the API interface,which solves the problem of ship model resources sharing.Experimental results show that the platform runs smoothly and performs well in concurrent testing and stress testing,meeting the requirement of uninterrupted operation for 24 hours.