Research On Garbage Classification And Detection Method Based On Deep Learning

With the continuous improvement of residents’ consumption power,more and more garbage is produced in daily life.Garbage classification is an important basis of harmless garbage treatment and resource recovery,as well as an inevitable trend of social development.At present,garbage classification methods mainly rely on manual classification in garbage collection stage,which is difficult to achieve satisfactory results in consistency,stability and sanitary conditions.Therefore,using artificial intelligence technology to realize automatic garbage classification has important academic value and practical significance.The existing garbage classification and detection methods based on deep learning have two question.(1)As garbage may be discarded in any scene,the collected images often contain irrelevant background information,which will distract the attention of the network,thus greatly reducing the accuracy of garbage identification.(2)In the case of dim image and small garbage target,the existing detection methods will have the problem of false recognition and missing detection.Aiming at the influence of background in question 1,this paper proposes a single-objective garbage classification method based on attention mechanism,which can make the attention focus more on garbage rather than background information.This method mainly consists of two parts: Frequency Mixed Attention Mechanism(FMAM)and lightweight classification network.In the pre-processing stage of feature information,FMAM uses DHT transformation to replace pooling in CBAM,which converts the feature information into multiple frequency components,then uses channel attention to obtain the frequency component of interest,and finally obtains the position of the frequency component of interest through spatial attention.Separable convolution is introduced into the classification network,which greatly reduces the parameters required by the network while ensuring high classification accuracy.To solve the problem of incorrect identification and missing Detection in question 2,a multi-scale Garbage Detection with Brightness Enhancement(MGDBE)method is proposed in this paper.This method consists of two modules: brightness enhancement module and multi-scale garbage detection module.In the brightness enhancement module,the long jump connection is used to enhance the connection of each convolutional layer,and the brightness adaptive enhancement is realized.In the main network of multi-scale garbage detection module,inverse residual block is introduced to reduce the information loss caused by downsampling.In the process of multi-scale feature fusion,features of different levels are fused together in the form of dense connection,which improves the learning ability of fine garbage features.In this paper,sufficient experiments are carried out around the above problems.Experimental results show that the proposed garbage classification method can reduce the background influence of garbage images and greatly reduce the network parameters compared with the existing methods.The proposed garbage detection method has higher detection accuracy than the existing method even when the image is dim and the garbage target is small.