An Identifying Campus Garbage Application Based on Optimized Convolutional Neural Networks

This paper proposes to present an effective and practical mobile garbage identification application named GCNet (Garbage Classification Network), which is capable of classifying different types of campus garbage and estimating the results using convolutional neural networks. The GCNet consists of three major parts. First, a data enhancement based on image stitching is designed to enrich data sets and enhance network robustness. Then we propose a spatial attention for reducing model structure and detection accuracy enhancement. Deformable convolution, the third component, is used to solve the problem of loss of sampling details or burr of normal convolution. Furthermore, we deployed the GCNet model to the mobile terminal and designed a mobile terminal intelligent campus garbage identification application. The experimental results show that the proposed GCNet performs well on campus garbage detection with 89.8% mean average precision (mAP), which outperforms the state-of-the-art methods. The running time on our application could achieve 0.083s per image, meeting the real-time detection. The proposed method is effective and applicable for accurate and real-time campus garbage detection.