Research On Gas Target Detection Method Based On Multi-beam Sonar Image

With the large-scale exploitation of marine oil and gas resources,the use of pipelines to transport natural gas is the transportation method used by most mining platforms.The transportation pipelines will be affected by corrosion and bad weather,resulting in natural gas leakage,and natural gas leakage in the marine environment Will cause adverse effects such as environmental pollution,so the detection of gas leakage is very important.In this paper,the sonar image data collected based on the AUV equipped with multi-beam sounding sonar is used as the data set to study the gas target detection method based on image processing.This paper organizes and establishes the data set collected by the multi-beam sounding sonar,analyzes various situations of the data set,divides the training set and the test set,and according to the distribution ratio of the image in the overall data under various conditions Assign the images of the training set and test set.Aiming at the problem of excessive sidelobe interference,OS-CFAR method is used to suppress sidelobe interference,so that the characteristics of the gas target are more complete.For the problem of a small number of data sets,the WGANs(Wasserstein Distance Generative Adversarial Network)method is used to generate and evaluate the gas samples and then put them into the background image of the gas-free target.Research on the data set based on artificially designed feature detection methods and deep learning detection methods.In the part based on artificial design features,the two features of HOG(Histogram of Directional Gradient)and LBP(Local Binary Pattern)are fused according to the feature design of the gas target,and the HOG feature is reduced in dimensionality,feature extraction is performed on the sample and the SVM is trained Classifier.For the detection of the entire image,the image pyramid model is used to generate multi-scale images to adapt to different sizes of gas targets,the sliding window is used to traverse the entire image with a certain step length,and the trained classifier is used to classify the images in the window.The resulting result boxes are classified according to their position relations and the detection boxes of the same category are merged.Based on the detection method of deep learning,the region-based target detection algorithm Faster R-CNN model is selected for research.Aiming at the problem of under-fitting the model due to less training set data,the training set data is expanded with the generated data,and Use the expanded data for training;the feature extraction network part uses the FPN(scale feature pyramid)idea to improve the resnet network,which improves the model’s feature extraction ability for targets of different scales,and realizes the information fusion between deep and shallow layers;The gas targets are sparsely distributed,and the candidate region extraction part uses the GA-RPN(Guided Anchoring-Region Proposal Network)method to replace the original RPN(candidate region extraction)network to generate high-quality candidate frames to improve the accuracy and detection of detection The precision of the box.After comparing the two methods of manual feature extraction and deep learning,the two types of detection methods have obtained better detection results,and the detection accuracy in different situations is relatively balanced.Although the recall rate and detection frame accuracy of feature detection methods based on manual design are not as high as those based on deep learning methods,they do not require a large number of training samples.Even if the data set is not expanded,better detection results can be obtained.The target detection algorithm does not need to design a more tedious manual feature extraction method for the data.