SOFTWARE ARCHITECTURE OF INTELLIGENT OBJECT-ORIENTED DATA FILTERING SYSTEM FOR NEURAL NETWORK CLASSIFICATION OF HOUSEHOLD WASTE USING CLOUD TECHNOLOGIES

Abstract

The paper proposes a software architecture for an intelligent object-oriented data filtering system for neural network classification of household waste using managed cloud computing nodes. The relevance is due to the need for stable operation of computer vision in realistic scenes, where image quality, background noise, glare, and class imbalance significantly reduce the reliability of decision-making. Unlike approaches in which data preparation is a one-time step before training, the proposed system integrates quality control directly into the training and inference cycle. The filtering module filters out low-informative images based on sharpness, contrast, and exposure balance with control over the preservation of class representation; the cleaned sample is used for further training of the basic architecture. The implementation is built on MobileNetV3-Small with feature transfer and replacement of the classification head with 30 classes; execution, artifact logging, and data storage are provided on Google Colab sessions with GPU and Google Drive/Kaggle storage, which guarantees reproducibility and portability of experiments. For user interaction, an inference web interface on Gradio has been created, which provides image loading, model configuration selection, and metrics viewing.

The experimental evaluation was conducted on the Recyclable and Household Waste Classification Dataset, which contains 15 thousand 256×256 images in 30 categories with controlled and real scenes. The basic configuration on the “raw” sample demonstrates consistent integral indicators, however, the inclusion of quality-oriented filtering gave subject-specific improvements for sensitive classes: in particular, for paper_cups, accuracy and completeness significantly increased, positive changes were recorded for steel_food_cans, clothing, and magazines, while for shiny and low-texture categories, false positives were mainly reduced. The results obtained confirm that the improvement of classification accuracy is achieved primarily through the optimization of input data and the discipline of the experiment in the cloud environment, without complicating the model architecture. The practical value lies in the creation of a reproducible methodological chain from controlled filtration to productive application, suitable for implementation on sorting lines and in the infrastructure of the circular economy.