Research And Application Of Sea Surface Floating Garbage Detection Based On Deep Learning

With the continuous development of China’s coastal industry and the continuous deepening of marine development,the pollution of sea surface floating garbage has become an urgent problem to be solved in China.Floating garbage on the sea not only causes visual pollution,affects coastal tourism,but also may damage the regional ecological chain.In recent years,AI technology based on AI video processing has been widely used in medical,remote sensing and other fields,but there is still less research on real-time identification of floating garbage on the sea surface using video monitoring.Using intelligent cameras to monitor the floating garbage on the sea surface can grasp the pollution status of floating garbage on the sea surface of the target sea area in real time,and provide support for scientific treatment of garbage pollution in the target sea area.Based on this,this paper proposes a method of detecting and applying the floating garbage on the sea surface based on deep learning.This method uses the floating garbage data set taken in different areas,and combines the Hikvision Embedded Open Platform(Hikvision Embedded Open Platform)intelligent camera head to carry out research on automatic identification and classification of floating garbage on the sea surface,and makes statistics on the quantity of the detected floating garbage on the sea surface,In addition,a method for calculating the size of floating garbage on the sea is also proposed.The research results can provide some reference for the treatment of floating garbage on the sea in China.The main work of this paper is as follows:(1)Use 3D-CNN model to determine whether there is floating garbage in the target sea area.Manually annotate the video data taken in different regions between May 2021 and August 2022,with 0 and 1 representing no target and target respectively,classify all the annotated video data,and finally obtain 432 videos,including 1678 videos with single background.Design a 3D convolution neural network model,including 9Conv3 d layers,8 Batch Norm3 d layers,10 PRe LU layers,1 FC layer and 1Batch Norm1 d layer.Put the labeled single background data set and complex background data set into the model for training,and test the performance of the model with the data not participating in the training.The final experimental results show that in the training set,the precision,recall and F1 values of the model under single background are 98.88%,99.13% and 98.88% respectively,while the precision,recall and F1 values of the model under complex background are 96.2%,95% and 95.2%respectively;In the test set,the values of the three evaluation indicators of the model under single and complex background are also above 98% and 94%.It shows that the3D-CNN model designed in this paper has high accuracy of judgment and can be applied in practice.(2)The 2D-CNN model is used to detect the type and location of floating garbage on the sea,and calculate the quantity and size of garbage.The images taken in different regions between May 2021 and August 2022 are labeled with data,which are divided into three categories: "Driftwood","plain" and "other".Then the YOLOV4 model is optimized,and the original feature extraction network is modified to Mobile Net V3network;In addition,SENet is added to the eighth layer of Mobile Net V3 network,and the Re LU in the third and fifth layers is modified to HS activation function,so as to improve the sensitivity of the model to the target area and avoid the neuron death problem in the Re LU activation function;Finally,K-mean++algorithm is used to optimize the prior frame of YOLOV4 model.The final results show that compared with the traditional two-stage model and the YOLO model of the same series,the optimized m AP value is 91.75%,which is the highest of all the comparison models,and the actual detection effect is also the best of all the models;In addition,in order to explore the impact of each module on the optimized model,the ablation experiment was designed to calculate the changes of the model parameters and calculation amount after each module was added.At the same time,combined with the target category and location detected by the model,a method to determine whether the target detected in the two frames before and after the video is the same target and a method to calculate the size of the target object are designed respectively.Under the condition that the target category detected in the two frames before and after the video is the same,whether the target object is the same object is determined by calculating the ratio of the area of the target object detected in the two frames before and after the video,the size of coincidence,and the size of color difference,Thus calculating the number of detected targets in a period of time;The method to calculate the size of the target object is to use geometric method to calculate the size of the detected target object according to the angle between the sea antenna and the camera and the distance from the camera to the water surface.(3)Use HEOP to complete the practical application of the deep learning model,use the HIKFlow platform to convert the trained deep learning model into a binary bin file,then use the converted bin file to complete the preparation of the reasoning base Demo file again through the HIKFlow platform,and finally use the packaging function of the HIKFlow platform to complete the app conversion of the deep learning model and embed it into the HEOP intelligent camera.The final field test results show that embedding the trained deep learning model into the HEOP intelligent camera can accurately detect the floating garbage in the target area.The innovation of this study is mainly reflected in the following three points:(1)A 3D-CNN model is proposed to determine whether there is floating garbage in the target area,which can provide more accurate judgment results for relevant staff.After the target is determined,the 2D-CNN model is used to detect the category and position of the target to reduce the interference of sea spray and light shadow on the2D-CNN model detection.(2)The original YOLOV4 model was optimized to make the m AP value of the model for floating garbage detection reach 91.75%;At the same time,the method of calculating the number of targets in the detection area and the method of calculating the size of targets are also designed.To provide more reference information for accurately understanding and detecting the garbage pollution in the sea area.(3)With the help of HIKFlow platform,the trained deep learning model is embedded into the Hikvision HEOP intelligent camera for practical application,realizing edge computing of data,and directly returning text information such as the category and location of the detected target,avoiding the problem of expensive maritime communication and large communication costs caused by the direct return of images from traditional cameras.