A Comparative Study of Azure Custom Vision Versus Google Vision API Integrated into AI Custom Models Using Object Classification for Residential Waste

Abstract

The residential separate collection of waste is the first stage in waste recyclability for sustainable development. The paper focuses on designing and implementing a low-cost residential automatic waste sorting bin (RBin) for recycling, alleviating the user’s classification burden. Next, an analysis of two object identification and classification models was conducted to sort materials into the categories of cardboard, glass, plastic, and metal. A major challenge in sorting classification is distinguishing between glass and plastic due to their similar visual characteristics. The research assesses the performance of the Azure Custom Vision Service (ACVS) model, which achieves high accuracy on training data but underperforms in real-time applications, with an accuracy of 95.13%. In contrast, the second model, the Custom Waste Sorting Model (CWSM), demonstrates high accuracy (96.25%) during training and proves to be effective in real-time applications. The CWSM uses a two-tier approach, first identifying the object descriptively using the Google Vision API Service (GVAS) model, followed by classification through the CWSM, a predicate-based custom model. The CWSM employs the LbfgsMaximumEntropyMulti algorithm and a dataset of 1000 records for training, divided equally across the categories. This study proposes an innovative evaluation metric, the Weighted Classification Confidence Score (WCCS). The results show that the CWSM outperforms ACVS in real-world testing, achieving a real accuracy of 99.75% after applying the WCCS. The paper explores the importance of customized models over pre-implemented services when the model uses characteristics and not pixel-by-pixel examination.

Keywords

• AI models
• object classification
• image processing
• real-time applications
• automatic bin
• recycling