Research On Garbage Detection Method Based On Convolutional Neural Networ

With the increase of garbage production,garbage image processing technology based on deep learning has become the focus of research.At present,the balance and optimization among detection accuracy,detection speed and model size are not fully considered when processing single-object garbage image classification and multi-object garbage image detection.In order to deal with these challenges,this paper proposes a garbage detection method based on convolutional neural network,which realizes the classification of single object garbage images and the detection of multi-object garbage images.The work of this paper is mainly divided into two parts:(1)Existing single-object garbage image classification algorithms do not fully consider the balance between detection accuracy and model size when improving detection speed,this paper proposes a single object garbage image classification algorithm based on LW-GCNet,based on lightweight garbage classification network.The convolution blocks in the feature extraction module of the algorithm all adopt the depth-separable convolution mode.Squeeze and Excitation(SE)modules are introduced in the first to third convolution modules.Feature fusion module extracts the feature output of the fourth and fifth convolution modules to carry out feature fusion of different sizes.In addition,the integrated features are processed by the adaptive maximum pooling layer and the global average pooling layer respectively,and the global features of the garbage image are extracted.Finally,the LW-GCNet performance was ablated and compared with garbage data set GRAB125.Under the premise of detection speed of 100 frames per second,the accuracy rate of LW-GCNet was 75.17%,and the number of parameters was 3.15 M.(2)Due to the large scale,poor real-time performance and low accuracy of the multi-target garbage image detection method,in this paper,a multi-objective garbage image detection algorithm is proposed based on an improved single-stage garbage detection network.In the shallow feature extraction module,depth-separable convolution is used for downsampling.In the deep feature extraction module,ECM convolutional module is used to extract features after high-low dimension conversion of feature images to reduce feature loss and parameters.A residual connection is introduced in the first to seventh convolution module,which makes the network still have generalization ability after deepening.Before the fifth,seventh,eighth and ninth convolution blocks,the space pyramid attention SPA module is introduced to enhance the representation ability of the original features of garbage images and significantly improve the accuracy of garbage detection.Finally,relevant experiments were carried out on 10000 garbage images of garbage classification and detection data set GCDD.The number of parameters of SP-SSD was 3.23 M,the number of m AP was 89.80%,and the detection speed was 93 frames per second.