ISSN :2582-9793

Outliers resistant image classification by anomaly detection

Original Research (Published On: 13-Mar-2025 )
DOI : https://doi.org/10.54364/AAIML.2025.51191

Soldatov Aleksei Valerievich, Anton Sergeev, Victor Minchenkov, Yaroslav Mazikov and Vasiliy Kakurin

Adv. Artif. Intell. Mach. Learn., 5 (1):3344-3355

1. Soldatov Aleksei Valerievich: National Research University Higher School of Economics

2. Anton Sergeev: National Research University Higher School of Economics

3. Victor Minchenkov: National Research University Higher School of Economics

4. Yaroslav Mazikov: National Research University Higher School of Economics

5. Vasiliy Kakurin: National Research University Higher School of Economics

Download PDF Here Citation Info via Semantic Scholar

DOI: 10.54364/AAIML.2025.51191

Article History: Received on: 07-Dec-24, Accepted on: 06-Mar-25, Published on: 13-Mar-25

Corresponding Author: Soldatov Aleksei Valerievich

Email: Soldatovalex34@gmail.com

Citation: Anton Sergeev, et al. Outliers Resistant Image Classification by Anomaly Detection. Advances in Artificial Intelligence and Machine Learning. 2025;1(1):191.


Abstract

    

The automatic monitoring of manual assembly processes in production settings increasingly relies on advanced technologies, including computer vision models. These models are designed to detect and classify events such as the presence of components in an assembly area and the connection of these components. However, a significant challenge for detection and classification algorithms is their vulnerability to variations in environmental conditions and their unpredictable behavior when encountering objects that are not present in the training dataset. Due to the impracticality of including all potential objects into the training sample, alternative solutions are needed. This study introduces a model that combines classification with anomaly detection by leveraging metric learning to create vector representations of images in a multidimensional space, followed by classification using cross-entropy. A dataset of over 327,000 images has been prepared for this purpose. Comprehensive experiments were conducted using various computer vision models, and the performance of each approach was systematically evaluated and compared.

Statistics

   Article View: 1591
   PDF Downloaded: 18