Garbage classification algorithm based on improved lightweight network shufflenetV2

Abstract

Aiming at the problem that the current deep learning garbage classification model cannot take into account high accuracy, small model size and high real-time performance, a garbage classification algorithm based on the improved lightweight network Shufflenet V2 is proposed. By improving the basic unit of Shufflenet V2 and introducing the attention mechanism module of CBAM, the feature extraction ability of the network was enhanced. Reduce the number of units used in each stage layer in the network structure, reduce the amount of parameters and calculation of the model, and at the same time, avoid excessively deep networks to extract irrelevant information, resulting in loss of accuracy. Replace the ReLU activation function with the Leakyrelu activation function to increase the richness of the extracted feature information. Use label smoothing loss function to reduce the negative impact due to sample class imbalance. The experimental results show that the accuracy of the algorithm on the self-built dataset is 81.26%. The parameter quantity of the model is about 0.917M, and the calculation quantity is about 92.75MFlops and 182. 32M MAdd. The accuracy of the algorithm is better than Resnet101, and the parameter quantity is only 1/44 of Resnet101, which provides a reference for the deployment and application of garbage classification and identification method in resource constrained devices such as mobile terminals.