基于GINet的垃圾分类检测网络
2021-07-11吕泽正祁翔
吕泽正 祁翔



摘 要:为了更好地服务于城市垃圾分类,提高垃圾分类前端收集的工作效率,本文提出了一种基于GINet的智能分类垃圾网络。首先在Kaggle数据集和华为垃圾分类公开数据集的基础上进行了人工标注,并建立了垃圾分类的训练数据集。其次,为了提高模型的泛化能力,扩充训练样本,设计了一种具有针对性的多背景图像增强方法。最后,为了解决垃圾分类数据集中某些同类垃圾之间的尺寸、颜色差异巨大,以VGG-16为主干特征提取网络,构建了一个融合多特征提取与注意力机制的垃圾识别网络(Garbage Identification Net,GINet)。仿真實验表明,该算法在复杂环境下拥有良好的鲁棒性和稳定性,检测准确率可达到945%,很好地满足了工业场景下垃圾检测的准确性。
关键词: 垃圾分类;深度学习;GINet
文章编号: 2095-2163(2021)01-0152-05 中图分类号:TP391 文献标志码:A
【Abstract】In order to better serve the classification of urban garbage and improve the efficiency of front-end collection of garbage classification, this paper proposes an intelligent garbage classification network based on GINet. Firstly, manual annotation is carried out on the basis of Kaggle and Huawei's public garbage classification data set, and a training data set for garbage classification is established. Second, in order to improve the generalization ability of the model and expand the training samples, a targeted multi-background image enhancement method is designed. Finally, in order to solve the huge difference in size and color between some similar garbage in the garbage classification data set, VGG-16 is used as the main feature extraction network, and a garbage identification network that combines multi-feature extraction and attention mechanisms is constructed named Garbage Identification Net (GINet). The simulation experiment shows that the algorithm has good robustness and stability in complex environments, the detection time is only 20 ms, and the accuracy can reach 94.5%, which satisfies the high efficiency detection of garbage targets in industrial scenarios.
【Key words】garbage classification; deep learning; GINet
0 引 言
近年来,中国国内城市垃圾产生量每年都在不断增长,因此,垃圾的科学分类与处理就显得尤为重要。一般情况下,垃圾处理主要是通过焚烧与填埋方式,但是这种做法却会导致土壤、空气和水资源受到严重污染,存在很大的弊端。据统计可知,城市生活垃圾主要分为干垃圾、湿垃圾、可回收垃圾和有毒有害垃圾[1]。就目前来说,国内关于垃圾回收大多采用了人工流水线处理,但在处理过程中则暴露出效率低、成本高且对工人身体健康影响较大的问题与不足。随着工业机器人技术的快速发展,机器人代替人工作业已在多个领域开始落地实施,因此基于机器人的垃圾分类回收作业已经成为该行业的未来发展趋势。……
