Research On Detection And Segmentation Of AUV Hydrothermal Deposits

Marine resources account for a large proportion in the earth’s resource composition system due to it’s vast surface area.A great deal of exploration has been carried out around the world to acquire marine resources.Because of the extreme ecological environment and large amount of seabed mineral resources,the deep-sea hydrothermal area is of high scientific and economic value,so the exploration of deep-sea hydrothermal area is one of the important research directions in the field of Marine science.In the case that the established sea area is known to have submarine hydrothermal area,the scientific research work needs to use the underwater vehicle to carry on the detailed exploration of hydrothermal area,at this time,the construction of the hydrothermal area environmental map,as well as the identification and detection of hydrothermal accumulation body in the hydrothermal area is extremely important.Due to the limited endurance capability of underwater vehicle in deep sea,object appearing in the map is detected based on deep learning to determine whether it is a hydrothermal accumulation body,so as to determine whether the exploration task is carried out or not.If it is a hydrothermal accumulation body,detour detection is carried out;if it is not a accumulation body target,no detection is carried out,so as to improve the hydrothermal area detection efficiency of underwater vehicle.Traditional deposit of target detection box can not describe the shape of deposit and super pixel division can obtain the shape of the contour feature,the split of pixel features are marked in the map,can obtain more accurate map of hydrothermal area environment,auxiliary underwater robot effectively detour and obstacle avoidance,provide guarantee for the safe navigation of underwater robot.Aiming at the deep-sea hydrothermal region detection problem,this paper studies the deep-learning-based hydrothermal accumulation target detection method and the superpixel-based hydrothermal accumulation segmentation method.The main contents are as follows:The research background and feasibility of underwater robot exploration of hydrothermal area were analyzed,and the underwater target detection method and image super-pixel segmentation method based on deep learning were summarized,so as to provide a foundation for subsequent research on the algorithm.The structural principles of convolutional neural network and classical network are illustrated,the characteristics of commonly used target detection algorithms are analyzed,and the theoretical basis is provided for the improvement of subsequent detection algorithms.Meanwhile,the target detection data set and segmentation data set of submarine hydrothermal accumulation body are established.Aiming at the problem of similar target and background as well as detection accuracy of few samples,the SSD algorithm was improved by adding feature fusion module,and the DI-SSD target detection algorithm for hydrothermal accumulation was proposed.On the basis of keeping the lightweight network,the feature fusion algorithm of cascade convolution and the enhancement of receptive field is added.Meanwhile,the feature characteristics of the upper layer and the lower layer are combined to enhance the ability to acquire features of different levels and improve the network detection performance.In the training stage,the transfer learning and the three-stage iterative method are introduced to update the parameters to improve the detection effect of the small data set accumulation object.The performance of DI-SSD algorithm is verified on the target detection data set of submarine hydrothermal accumulation.The results show that the algorithm has higher detection accuracy.To improve the precision of the navigation map,and meet the requirements of real time,based on the SLIC algorithm,through the location coordinates and color information value is stored as a plastic,initial clustering center edge when L channel value,only to calculate the relative distance clustering on the results don’t open square,and near to the combined methods of edge pixel block,under the premise that segmentation accuracy is guaranteed to speed up the algorithm segmentation,improve the edge pixel block and coincidence degree,the edge of deposit by hydrothermal deposit divided datasets compare the performance before and after improvement,prove the improved R-SLIC algorithm effectively improves the segmentation of the target of hydrothermal deposit speed and quality of segmentation.The underwater accumulation body was detected by underwater robot.Build the test environment for detecting algorithm validation hydrothermal deposit,the use of underwater robot with optical camera,according to the scheme of probe test set two paths,hydrothermal deposit collected video image,the image detection and segmentation task,segmentation algorithm for detection and super pixels in the hydrothermal environment validity is verified.