Research On Ship Black Smoke Detection Method Based On Deep Learning Multi-Feature Fusion

Ship black smoke contains a lot of sulfide,PM2.5 and other harmful substances,which has become one of the important sources of marine and port pollution.However,at present,the management mechanism of real-time detection and graded quantification of ship black smoke has not been realized,and the management of ship black smoke pollution does not have a theoretical basis.Due to the rapid development of artificial intelligence in the field of vision,target detection technology provides a modern solution for ship black smoke management.Due to the rapid development of artificial intelligence in the field of vision,target detection technology provides a modern solution for ship black smoke management.This paper conducts research around the problem of ship black smoke identification,focusing on real-time detection of ship black smoke,on which a quantitative evaluation method of ship black smoke is proposed to provide reference for the management of ship black smoke pollution.The main research work of this paper is as follows:(1)To address the problem of lack of public data set for ship black smoke target detection.Video resources of ship black smoke in real scenes are collected and combined with data from web resources to construct a set of ship black smoke target detection datasets with rich background and variable scenes,and some of the datasets are atomized by standard optical model considering environmental interference factors.According to the application background,the experimental purpose and other factors of the target detection algorithm to YOLO,SSD as the representative of the one-stage algorithm and Faster-RCNN two-stage algorithm was carried out experimental analysis.The results show that YOLOv5s algorithm is more suitable for the detection task in this paper.(2)Research on ship black smoke detection algorithm based on optimized YOLOv5s multi-feature fusion.Firstly,a CBAM convolutional attention mechanism is added to the YOLOv5s network structure to improve the network’s attention to the black smoke region of the ship and suppress the interference information;secondly,based on the idea of Bidirectional weighted feature pyramid network Bi FPN,a lightweight network structure Tiny-Bi FPN is designed to realize the feature fusion path with multiple scheduling;thirdly,an adaptive spatial feature fusion mechanism is introduced ASFF to suppress the effect of feature scale differences;finally,EIo U＿Loss is used as the localization loss function to improve the regression accuracy and convergence speed of the model.The optimized YOLOv5s algorithm improves 3.8%,5.7% and 4.5% in detection accuracy,recall rate and average accuracy m AP 0.5,respectively,compared with the original model.(3)Research on the evaluation method of ship black smoke based on Ringelmann blackness.Firstly,the K-means clustering algorithm was optimized to determine the clustering centers by mining color information through color histograms,and the similarity of samples was calculated using the Marxian distance;secondly,the improved segmentation algorithm was compared with Otsu,region growing and other segmentation algorithms,and MSE and PSNP both performed optimally,and the segmentation efficiency was improved by 30%;finally,the segmented background was used as the reference system to classify the blackness level according to the Ringelmann blackness method,and the ship blackness level was estimated by the weighted average grayness ratio of the effective black smoke area to the background area.The ship black smoke is detected by the YOLOv5s-CMBI algorithm with multifeature fusion,and a ship black smoke evaluation method based on the Ringelmann blackness is proposed to quantify the blackness level.The experimental results show that the accuracy of the algorithm for ship black smoke blackness evaluation can reach 92.1%,which provides a theoretical basis for the process of ship black smoke pollution control.