Garbage Image Classification Based on Improved ShuffleNet Fusion Algorithm

• Published in: 2024 IEEE International Conference on Cognitive Computing and Complex Data (ICCD) pp. 57 - 62
• Main Authors: Shang, Kaikai, Xiao, Shaozhang, Lin, Ken, Xiao, Kangliang, Xia, Yin
• Format: Conference Proceeding
• Language: English
• Published: IEEE 28.09.2024
• Subjects: Accuracy; Computational modeling; Convolution; Data mining; Data models; Depthwise Separable Convolution; Feature extraction; Image classification; Information retrieval; Merging; MobileViT; Neural Network; Periodic structures; ShuffleNet V2
• DOI: 10.1109/ICCD62811.2024.10843513

Aims to address the problems of low accuracy of existing garbage classification and low classification ac curacy of models, large model size and difficulty of deployment on portable devices. A lightweight garbage image classification model is proposed by merging the improved lightweight model ShuffleNet V2 and MobileViT. Firstly, the structure of the ShuffleNet V2 model is adjusted and improved to enhance the cross-channel information interaction ability of the model, which improves the feature information interaction ability between the channels of the ShuffleNet V 2 model; secondly, the improved ShuffleNet V2 model is fused with the MobileViT model, the advantage of the MobileViT module in global feature information extraction is used to overcome the shortcomings of the ShuffleNet V2 network, which is difficult to extract global feature information. Comparative experiments are conducted on public datasets to evaluate the performance of the model. The experimental results show that the accuracy of the proposed SNViT model is significantly better than the D enseNet-121 model, and also has a better classification performance compared to EfficientNet V2, MobileNet V2 and other models, while the precision rate, recall rate and F1 value are also improved. The improved SNViT model improves the accuracy of the image classification task and achieves high classification performance, while the number of parameters is small and easy to train and deploy.