Image Classification in the Browser: a performance assessment

During the last decades, a lot of step forward are made in Artificial Intelligence field. Until recently, the Cloud Computing paradigm has allowed for increasingly complex and large models, but recently a paradigm shift has occurred and Edge Computing has taken over having an eye on issues such as privacy and portability. In the last years, JavaScript libraries have emerged, allowing to Deep learning to be brought in the browser. These libraries provide several benefits, in particular ensure the portability. WebAssembly is a low-level binary format that is designed to be executed by web browsers. It provides a way to run code in the browser that is closer to native machine code than JavaScript. This means that Deep Learning models built using JavaScript libraries can be deployed with WebAssembly, on a wide range of devices and platforms, making it easier to integrate Deep Learning into web applications. The solution presented, is a static web application that perform a classification task on emotional state of people in work environment. The use of static site ensure the privacy and the ONNX Runtime Web enables ONNX Deep learning models to run in the browser. Results are obtained with testing the web application of three devices with different hardware performance. Eight Convolutional Neural Networks with different depth and complexity are tested. The web application produces three outcomes: the latency time, the prediction outcome and its probability. This work proves that JavaScript libraries are a good solution to overcome the issue of portability and, in particular, put the focus on the performance obtained with web application.