A LIGHTWEIGHT MODEL FOR PAVEMENT GARBAGE CLASSIFICATION BASED ON DEEP LEARNING

Abstract

In response to the requirement of urban environmental health to classify and recycle citizens' garbage, a lightweight model for pavement garbage classification based on deep learning is proposed considering the powerful performance of computer convolutional neural networks in image classification. In order to reduce the scale of the network model to run better on embedded devices, on the basis of the original convolutional neural network framework, the ghost module is reasonably embedded into the lightweight model for pavement garbage classification and the ordinary convolution layers in the path aggregation network (PANet) are replaced by the depthwise separable convolution layers. In order to improve accuracy and real-time performance, the attention mechanism squeeze and excitation layer (SELayer) is embedded to fuse more spatial features. Then, a self-made dataset is built through internet search and manual photography. The self-made dataset contains seven types of common pavement garbage images. Through experimental demonstration, the proposed model can well locate the garbage and identify the types of garbage. And the proposed model has the advantage of small size. The proposed model also achieves 15.58 frames per second (FPS) real-time performance, and the size of the model is only 5.18 mbyte (MB). The FPS increases by 27.5%. The model size is reduced by about two thirds.