Research On Garbage Classification And Detection Based On Deep Learning Network

As we all know,for the problems of various types,complex components and large area of garbage,it will lead to serious pollution of water,soil,atmosphere and so on.At the same time,the garbage will also destroy the ecology,leading to species extinction,global warming and other major world problems.If these wastes are not treated scientifically,they will not only pollute the environment,but also destroy the ecology.Therefore,the classification of garbage is very important.The urgent tasks are to reduce the pollution of garbage to the environment,carry out scientific and rational recycling of garbage,and reduce the damage of garbage to the environment.Therefore,it is obvious that recycling is very important in modern society.At present,in order to solve the problems of poor environment,heavy tasks and low efficiency of artificial garbage classification,a lightweight feature fusion single multi-box detector algorithm was proposed in this dissertation to realize intelligent garbage classification and recognition.Due to the small size of garbage and low resolution of garbage image,the algorithm was proposed in this dissertation has a lightweight and novel feature fusion module,and it can significantly improve the performance of garbage detection.This method can be fully applied in practical application,and we can apply it to the garbage identification and classification of intelligent trash cans or the process of garbage sorting in large garbage dumps,so as to reduce the burden of manual classification.This method can also play a significant role in the process of self-recovery and utilization in production equipment.Due to the above problems,this dissertation was proposed that based on the optimization of the existing deep learning network,and compared the advantages of the proposed model through experiments.The main research contents are as follows:(1)An improved Faster R-CNN algorithm is proposed to deal with the problem of garbage classification,to overcome the problem of manual preprocessing of garbage classification,and to complete the tasks of garbage detection and classification with better quality.Explore the impact of VGG16 network architecture and Res Net(Residual Network)architecture on the performance of garbage classification algorithms,so as to select a better network to help us complete the classification tasks.(2)The framework of feature fusion is redefined,and some improvements are made on this structure of the original pyramid.This new structure can fuse the semantic information of each layer together,so that the final detection result is closer to the label category.It is a very simple but compact method that does not take up too much computing space.By generating a new feature pyramid,the features of each drawing can be reintegrated.(3)Focal Loss function was proposed in this dissertation to replace the traditional loss function.The problems caused by the imbalance of categories that are often encountered by One-stage methods in recognition training.These detectors evaluate a number of candidate locations for each image,but only a few of them contain their target objects.However,through experiments,it is found that the Focal Loss function can smoothly deal with the class imbalance problems,which enables us to train all samples effectively.In the process of training,by reducing the weight of the samples that are easy to classify,this model will pay more attention to the samples that are difficult to classify.(4)By analyzing and optimizing the influence of NMS(Non-maximum Suppression)algorithm on garbage classification,it is found that this strategy is more suitable for obtaining better image representation.In addition,the sensitivity of the parameters of Soft-NMS is analyzed.The experimental results show that the parameters of this algorithm are in the range of 0.4-0.7,which can significantly improve the performance of garbage detection.(5)A corresponding garbage detection and classification which was proposed in this dissertation system is designed based on the algorithms.The system applies and tests the proposed algorithms.Input the image form the data source,and then users can select the corresponding algorithms to detect different images.Finally,the output of this detector will show the corresponding detection category and the detection accuracy.