Image Similarity Search for Industrial Objects with the Main Focus on Siamese Networks

Abstract

This thesis investigates the effectiveness of different image similarity search methods for
industrial object retrieval, focusing on Siamese networks. In industrial environments, visual
inspection tasks require accurate and efficient retrieval of similar images under real-world
conditions. The study compares classical methods with deep learning approaches, including
pre-trained CNNs, autoencoders, and Siamese architectures. An emphasis is placed on
sequential training strategies, where encoders pre-trained through autoencoders or classification
are fine-tuned within Siamese networks. Experiments are executed on the MNIST dataset and
a proprietary industrial dataset. Evaluation metrics include retrieval accuracy, computational
efficiency, and training time. Results indicate that pre-trained CNN-based methods provide
strong baseline performance. At the same time, sequentially trained models such as fine-tuned
Siamese networks with autoencoder-pre-trained encoders showed improved results on MNIST
but did not consistently outperform other methods on the industrial data. The findings offer
guidance for developing efficient image retrieval systems connected to industrial applications.

Date of Award	2025-Jun
Original language	English
Supervisor	Niklas Gador (Supervisor), Kamilla Klonowska (Assessor) & Magnus Johnsson (Examiner)