ALGORITHMS FOR IMPROVING THE ACCURACY OF NEURAL NETWORK CLASSIFICATION OF HOUSEHOLD WASTE USING CLOUD-MANAGED COMPUTING NODES

Abstract

The relevance of the work is due to the growth of household waste flows and the need for accurate recognition of material categories in realistic scenes with uneven lighting, background noise and class imbalance. The practical effectiveness of such systems is determined not only by the choice of architecture, but primarily by the controlled quality of data and the reproducibility of experiments in a standardized environment. The article proposes a quality-oriented pipeline in which the quality control module is integrated directly into the training cycle. Filtering by sharpness, contrast, exposure balance and background clutter forms a cleaned subsample for further training of the pre-trained model. Managed cloud computing nodes based on Google Colab sessions with access to graphics accelerators, fixed library versions and artifact logging are used, which ensures the stability of the software environment and the comparability of series.

The methodology is based on the MobileNetV3 Small architecture with ImageNet feature porting and replacing the classification head with a thirty-class problem statement. The Recyclable and Household Waste Classification Dataset with thirty categories, including paper, plastic, glass, and metal subclasses, as well as organic fractions, was used for experiments. The baseline evaluation on the raw sample yielded consistent results across metrics with an accuracy of 0.7703 and high areas under the ROC curves, indicating good resolution of probabilistic outputs and a reserve for stabilizing solutions in multi-class mode. Inclusion of filtering in the training cycle provided subject-specific improvements in classes prone to cross-validation errors due to gloss and weak texture. For paper_cups, an increase in accuracy of 13.13 percent, completeness of 10.69 percent, and integral F1 of 11.85 percent was recorded. Positive improvements were also obtained for steel_food_cans, clothing and magazines, where confusion with visually similar categories was reduced.

The results obtained confirm the feasibility of shifting the emphasis from the complexity of architectures to managed data quality and the discipline of experimentation in a cloud environment. The proposed integration increases the robustness of classification and creates a basis for the reliable implementation of computer vision in the recycling infrastructure and supports circular economy practices.