Research On Object Detection And Segmentation Of Underwater Pollutant Based On Improved Mask R-CNN

With the increasingly serious problem of marine trash pollution,the monitoring and management of marine trash is imminent.This paper firstly analyzes the research status of marine debris monitoring,demonstrates the advantages of image detection and segmentation methods in the field of marine trash monitoring,and proposes a marine trash detection and segmentation scheme based on the improved Mask R-CNN algorithm,which alleviates the problems about insufficient detection and segmentation accuracy of marine trash objects appeared in previous related research.Furtherly,Mask R-CNN was applied in the field of underwater acoustic object detection,and the underwater acoustic object detection scheme of this paper was proposed.First of all,for the problem of poor underwater optical image quality,the related research on underwater image enhancement algorithm is carried out.Firstly,based on the improvement of the existing IBLA enhancement algorithm,a relatively simple background light and depth estimation method is used to reduce the complexity of the algorithm,and the image enhancement algorithm of this paper is proposed,which effectively improves the image quality of deep-sea trash dataset Trash Can in this paper.Then,compared the enhancement effect and image quality evaluation index of the algorithm in this paper with several other classical algorithms,and verified the advantages of the algorithm in this paper compared with other classical algorithms.which also pave the way for improvement of the detection and segmentation accuracy of the Mask R-CNN model by using proposed enhancement algorithms.Then,this paper uses the Trash Can dataset to conduct a basic simulation of the Mask RCNN algorithm.Firstly,the training set and test set are divided for Trash Can,and the simulation scheme of the Mask R-CNN algorithm is designed;then the Mask R-CNN algorithm is simulated,and the training and inference process of the Mask R-CNN are analyzed,and basic Mask R-CNN model are obtained;and after model evaluation,the detection and segmentation accuracy of the basic model are: detection AP 54.96% and segmentation AP 47.43%.Obviously,the detection and segmentation accuracy of the basic model cannot well meet the needs of deep-sea garbage monitoring.Then,this paper adopts a variety of improvement strategies to optimize the Mask R-CNN model,which greatly improves the detection and segmentation accuracy of the model.Firstly,a variable learning rate strategy was introduced to optimize the model training process;then a better Res Next101+FPN backbone network was used to replace the Res Net50+FPN network of the basic model,which improved the representational power of the entire network;then the image enhancement algorithm in this paper was adopted The Trash Can image is preprocessed to improve the image quality of the dataset;and for the unbalanced distribution of sample categories and the small amount of data in the Trash Can dataset itself,this paper adopts data balance and data augmentation solutions to alleviate the problem.what’s more,the model is finally optimized using the cascaded Mask R-CNN strategy.Furtherly,it is verified by experiments that the above improvement strategy can effectively improve the detection and segmentation accuracy of the Mask R-CNN model.Furtherly,this paper adopts the method of ablation experiment to compare and analyze the model performance evaluation results of each improved strategy,and obtain the optimal Mask R-CNN model in this paper.After model evaluation,the detection and segmentation AP of the optimal Mask R-CNN model reached 67.58% and 52.65% respectively,which is a great improvement compared to the basic Mask R-CNN model,and its detection and segmentation rate also reaches 11 FPS,which can meet the accuracy and real-time requirements for detection and segmentation of deep-sea trash objects.Finally,this paper applies the improved Mask R-CNN algorithm to the field of underwater acoustic object detection,and improves the accuracy of underwater acoustic object detection.Firstly,according to the characteristics of the underwater acoustic image dataset,some simulation parameters are redesigned and the Mask R-CNN algorithm is simulated to verify the effectiveness of the ImageNet pre-training model for acoustic target detection,and the basic acoustic target detection model is obtained.Then,the parameters of the optimization strategy are redesigned,and the optimal acoustic target detection model in this paper is proposed.After model evaluation,the model reaches 53.67% detection AP and 5FPS detection rate,which can be applied to the underwater acoustic object detection scene that does not require high detection accuracy and real-time performance.