| Garbage classification is of great significance for resource reuse,reduction of land erosion and environmental pollution.In order to overcome the disadvantages of traditional garbage classification,such as bad working environment,high labor intensity and poor effect due to relying on manual sorting,the intelligent garbage classification has attracted wide attention.Deep learning technology is used for automatic recognition of massive garbage images to explore the recognition accuracy and robustness in the intelligent garbage classification.Aiming at the problem of low recognition accuracy of Standard Junk Image Classification Dataset set when using ResNeXt network model,a ResNeXt garbage image classification model SE-ResNeXt based on attention mechanism is proposed.Firstly,the performance of five network models including Mobile Net＿v2,Shuffle Net＿v1,Shuffle Net＿v2,ResNet and ResNeXt for garbage image recognition is compared and analyzed.The experimental results showed that the ResNeXt model has better performance,but the recognition accuracy needs to be further improved.To this end,the depthwise separable convolution strategy of ResNeXt model and the feature extraction mechanism of the SE module are combined to increase the dimension information,enhance the feature processing capability of the model.The experimental results showed that the proposed SE-ResNext garbage image classification model improved the recognition efficiency and can obtain a recognition accuracy of 95.81%in the Standard Junk Image Classification Dataset,and the recognition accuracy of different types of garbage image were: 97.34% for recyclable garbage images,97.30% for kitchen garbage images,92.52% for hazardous garbage images,86.76%for other garbage images.In order to further enhance the robustness and classification effect of SE-ResNext classification model,a ResNeXt garbage image classification model SE-ResNext-All combined with Cut Mix data enhancement and Label Smooth regularization is proposed.Firstly,Cut Mix data augmentation algorithm is used to process the Standard Junk Image Classification Dataset,and enhance the diversity of the data set.Secondly,the label smoothing algorithm is used to reduce the weight of the real sample label category in calculating the loss function and enhance the anti-noise ability and generalization ability of the model.The experimental results showed that the proposed SE-ResNeXt-all model can obtain 96.31% recognition accuracy in the Standard Junk Image Classification Dataset,and the recognition accuracy of different types of garbage images were: 98.12% for recyclable garbage images,97.81% for kitchen garbage images,94.83% for hazardous garbage images,87.13% for other garbage images.On the basis of previous studies,aiming at the problems of large amount of garbage and complex background in multi-target mixed garbage images,a multi-target segmentation detection model of garbage images based on Mask R-CNN is proposed.Deep Cross Network and trunk Network ResNet50 are combined to efficiently capture effective features and improve the accuracy of the model in multi-target segmentation and detection of garbage images.The experimental results showed that the proposed Mask R-CNN garbage image multi-target segmentation detection model fused with DCN network had a good recognition effect on mixed garbage images,reaching 79.85% m AP in huawei cloud garbage detection data set,2.23 percentage points higher than the benchmark network Mask R-CNN.The average accuracy of detecting different types of garbage images were: 91.71% for recyclable garbage,66.43% for kitchen waste,83.02% for hazardous garbage,and78.23% for other garbage. |
