Optimization Model for Garbage Image Classification Based on RepVGG

• Published in: 2024 IEEE International Conference on Cognitive Computing and Complex Data (ICCD) pp. 77 - 82
• Main Authors: Lin, Ken, Xiao, Shaozhang, Shang, Kaikai
• Format: Conference Proceeding
• Language: English
• Published: IEEE 28.09.2024
• Subjects: Accuracy; Adaptation models; Computational modeling; Convolution; Deep learning; depthwise separable convolution; DSC5RepVGG; Feature extraction; Image classification; Kernel; Periodic structures; RepVGG; Stacking; Training
• DOI: 10.1109/ICCD62811.2024.10843599

The application of deep learning to the classification of garbage images has been demonstrated to significantly enhance the accuracy and efficiency of classification through the extraction of features and the recognition of patterns. However, in environments with limited resources or computational power, traditional models are unable to effectively handle complex trash image classification tasks. Consequently, the deployment of lightweight models in garbage image classification h as g arnered increasing interest. In this paper, the structure of the RepVGG model is optimized by combining depthwise separable convolution. Adapting the convolutional layers of the original trunk structure to training branches and replacing the convolutional kernel of the trunk with a 5 × 5 convolutional kernel results in the trunk and branches being reparameterized to a single 5 × 5 convolutional layer after training. This design enhances the image feature extraction capability at multiple levels. The inference structure retains the simple stacking of the original convolutional and ReLU layers. In this paper, the improved model is named DSC5RepVGG. The experimental results demonstrate that DSC5RepVGG reduces the number of parameters by 60.97% in comparison to RepVGG, reduces the theoretical floating-point computation by 58.22%, and improves the Top-1 accuracy by 1.29% on Kaggle's Garbage Classification dataset.