摘要
固体废物在城市地区的积累正成为一个令人关注的问题，如果管理不当，将会造成环境污染，并可能危害人类健康。拥有先进的垃圾管理系统来管理各种废物非常重要。然而，回收任务通常需要大量的人工成本，而计算机视觉和深度学习（DL）技术可帮助自动检测和分类废物类型以进行回收任务。深度学习是可以解决视觉识别和分类的人工智能部分之一，而卷积神经网络（CNN）是目前称为图像识别的主要方法的DL体系结构之一。基于有效图像识别的自动分类机器人可以帮助减少回收任务的繁重工作。卷积神经网络（CNN）模型，例如ResNet，改进了传统的图像识别技术，并且是当前图像识别的主要方法。
本系统是使用ResNet卷积神经网络模型开发的，该模型是一种机器学习工具并充当提取器，用于将垃圾分类为不同的类型，例如玻璃、金属、纸张和塑料等。经过导入科学库 TensorFlow、sklearn来搭建神经网络数据模型，定义算法核心的 bottleneck残差学习单元搭建ResNet-50数据模型同时引入softmax、relu、sigmoid等激活函数。导入训练集，进行不断的监督学习训练搭建垃圾识别模型。该系统在垃圾图像数据集上进行了测试，并能够在数据集上达到87％的准确性。
关键词：图像识别；卷积神经网络；ResNet算法
Abstract
The accumulation of solid waste in urban areas is becoming a concern. If not managed properly, it will cause environmental pollution and may endanger human health. It is very important to have an advanced waste management system to manage all kinds of waste. However, recycling tasks usually require a lot of labor costs, and computer vision and deep learning (DL) technology can help automatically detect and classify waste types for recycling tasks. Deep learning is one of the artificial intelligence parts that can solve visual recognition and classification, and convolutional neural network (CNN) is one of the DL architectures currently called the main method of image recognition. Automatic classification robots based on effective image recognition can help reduce the heavy work of recycling tasks. Convolutional Neural Network (CNN) models, such as ResNet, improve traditional image recognition technology and are currently the main method of image recognition.
This system is developed using the ResNet convolutional neural network model, which is a machine learning tool and acts as an extractor to classify garbage into different types, such as glass, metal, paper, and plastic. After importing the scientific libraries TensorFlow and sklearn to build the neural network data model, the bottleneck residual learning unit that defines the core of the algorithm builds the ResNet-50 data model and introduces activation functions such as softmax, relu, and sigmoid. Import the training set and conduct continuous supervised learning training to build a garbage recognition model. The system has been tested on the garbage image data set and can reach 87% accuracy on the data set.
Key words: Image recognition; convolutional neural network; ResNet algorithm
