Research And Implementation Of Recyclable Garbage Classification System Based On Object Detection Network

With the continuous development of the global economy,the impact of marine debris on marine life,water quality and other marine resources has attracted wide attention from all countries.At present,the cleanup of seabed garbage mainly depends on the underwater salvage of divers.Due to the special physiological structure of the human body,diving salvage has disadvantages such as inability to operate for a long time,hidden danger to life,and inability to penetrate deep sea areas.Therefore,research and development of underwater robots with functions such as underwater environment perception,underwater automatic path cruising,and underwater fishing operations have high theoretical significance and practical value for solving the problem of marine garbage salvage.The classification and positioning of underwater targets is an important part of underwater environment perception,and the result will directly affect subsequent manipulator fishing operations.In the detection tasks of large-scale land image data sets with classification and positioning as the target,the classification and positioning detection results of the object detection algorithm based on deep learning are better.However,in the detection task of underwater image data sets,due to the harsh underwater imaging environment,underwater images often show blur,distortion,low contrast,etc.,and are limited by factors such as equipment and manpower.The cost of collection and labeling is high,and the data cannot be obtained on a large scale.Therefore,it is impossible to obtain high classification and positioning accuracy through the existing detection algorithms.Aiming at the detection task of classification and positioning of underwater recyclable garbage,the main research work and research contents of this thesis are as follows:1.Analyze the main reasons of underwater image degradation and imaging model,combined with a variety of existing image enhancement algorithms to construct a small sample data set for underwater image enhancement algorithm training,and based on the imaging model,propose Image enhancement algorithm based on deep learning.This method constructs a convolutional neural network through the imaging model to calculate the parameters in the imaging model,and realizes the color correction and detail enhancement of underwater images.2.Use an object detection algorithm targeting classification and positioning,in view of the relatively small amount of data and categories of the constructed underwater data set,in the feature extraction part,the structure of the densely connected network is modified,and the residual mechanism is introduced to construct A feature extraction network Res Dense Net based on the residual densely connected network.This network combines the advantages of residual networks and densely connected networks,and has certain advantages compared with existing feature extraction networks in terms of parameter amount,calculation amount,and feature extraction capabilities.In the feature fusion implementation part,an improved spatial pyramid pooling structure and an adaptive spatial feature fusion structure are introduced to strengthen the fusion of multi-scale features between different layers of the network,and enhance the detection performance of the object detection algorithm for the classification and positioning of targets of different sizes.In the method of generating a priori box,the Kmeans++ algorithm is introduced to solve the problem of unstable cluster center generation in the Kmeans algorithm.In the position loss of the loss function,it is proposed to use the position loss based on CIOU to replace the square difference loss of the original network.3.Construct an underwater object detection algorithm UWNet.In terms of classification and positioning,the detection performance of UWNet algorithm was verified through experiments.Design and implement an underwater garbage object detection system based on deep learning,targeting classification and positioning.The system integrates functions such as image enhancement,classification and positioning,model training,visualization,etc.,which facilitates the use of various functions of the system by the operator.