Research On Detection Algorithm Of Marine Floating Objects In Complex Environment

Ocean shipping is an important part of global trade,accounting for more than 80% of global cargo trade.However,due to frequent ocean activity,maritime accidents occur from time to time,which has brought serious negative effects on the country’s development,causing not only huge economic losses but also threatening the safety of crew members’ lives.In recent years,with the continuous improvement of computer computing power,deep learning has also accelerated its development.Applying deep learning-based object detection algorithms to marine floating object detection and maritime rescue can not only save manpower and material resources but also improve efficiency,playing an important role in ensuring maritime safety and maritime rescue.However,due to the complexity and variability of maritime weather,and the fact that the targets in maritime images captured by devices such as drones are mostly small,higher requirements are placed on detection algorithms,and existing object detection algorithms are difficult to achieve satisfactory results.To address these problems,this paper studies small target detection algorithms in complex environments and proposes two efficient detection algorithms considering the requirements of different devices.The research work is as follows:(1)In view of the limited number of currently available public datasets for small target detection in the maritime field,this paper collects maritime target images through actual shooting and online collection methods,manually annotates and mixes them with some maritime images extracted from public datasets.With this self-made maritime dataset,the algorithm’s performance can be effectively tested in actual maritime scenes.(2)Complex scenes such as foggy weather and nighttime that may occur in actual marine environments may interfere with the detection model’s accuracy.To solve this problem,this paper adds an image preprocessing module during the training process and inputs the preprocessed and original images together into the backbone network to improve the detection model’s robustness.This strategy can effectively improve the accuracy of the detection model.(3)Feature extraction and fusion are critical steps in object detection models and important means to improve detection model performance.Considering that targets in maritime images captured by devices such as drones are mostly small,this paper optimizes the feature extraction network to reduce the loss of fine-grained features.In addition,in the feature extraction process,as the network depth increases,the feature layer will extract more semantic information but lose positional information.The feature fusion network fuses feature layers of different levels,making the fused feature layer possess rich semantic and positional information.To adapt to maritime scenes,this paper improves the feature fusion network of the baseline model to further improve the model’s detection performance.The proposed improved YOLOv5 algorithm improves accuracy by about 7 percentage points.(4)For terminal devices such as drones,the limited internal storage space and chip computing power make common detection models difficult to ensure real-time performance.To solve this problem,this paper adopts a lightweight model construction method,adjusts the baseline model’s width to reduce the parameter volume,and further compresses the model by combining knowledge distillation,channel pruning,and other technologies.To reduce the accuracy loss caused by lightweight,an attention mechanism is added to the lightweight model.The Slim-YOLOv5 model proposed in this paper based on channel pruning has a parameter volume of only 8.7% of YOLOv5 s but has higher detection accuracy.This lightweight model can ensure real-time operation on terminal devices and has high practical value.