Abstract. Automated object recognition systems are widely used in industry, robotics, and environmental monitoring, with intelligent waste sorting being one of the most promising application areas. Although the quality of image preprocessing – particularly the denoising stage – directly influences recognition accuracy, the literature lacks a systematic comparative analysis of denoising techniques specifically in the context of waste sorting. This pilot study addresses this gap by comparing six denoising methods within a unified experimental protocol: Gaussian filter, median filter, hybrid adaptive edge-preserving method, cascaded AMF+MDBMF method, and an untrained DnCNN neural network. The experimental dataset included 15 representative color images of recyclable objects from three classes (plastic, glass, metal) selected from the Kaggle TrashNet dataset, distorted by three types of artificially added noise: Gaussian noise (σ = 25), salt-and-pepper impulse noise (p = 5%), and Poisson noise. Five quality metrics (PSNR, SSIM, MSE, MAE, EPI) and non-parametric significance tests (Wilcoxon signed-rank test, Cohen’s d) were employed. In addition, the computational complexity of each method was measured. Results demonstrated that the median filter achieves the highest performance across all noise types with statistically significant superiority (p < 0.01). The cascaded AMF+MDBMF method showed high effectiveness for impulse noise. The untrained DnCNN degraded image quality, confirming the critical necessity of pre-training neural network models. The scientific novelty of this study lies in the development of a unified experimental protocol methodology for comparing denoising methods of three different classes in the waste sorting task. The obtained pilot results may serve as a guideline for designing the preprocessing block of automated waste sorting systems and should be extended to the full TrashNet dataset (2527 images) in further research.

Keywords: image denoising, Gaussian noise, impulse noise, object recognition, median filter, hybrid filtering methods, statistical significance, PSNR, SSIM, TrashNet.