Impact of Image Enhancement on Multi-Object Tracking in Underwater Scenario

Abstract

This thesis examines the impact of image enhancement techniques on multi-objecttracking (MOT) performance using four deep learning models: Long Short-TermMemory (LSTM), Vision Transformer, Siamese Network, and Convolutional NeuralNetwork (CNN). The objective is to assess the effectiveness of these modelsin handling challenging visual conditions and explore the benefits of image preprocessingtechniques for improving tracking accuracy.The study utilizes various image enhancement approaches, including denoising,deblurring, and super-resolution. Each deep learning model is implementedand trained on a large-scale dataset specifically designed for multi-object tracking.Performance evaluation is conducted on benchmark datasets, comparing thetracking accuracy of the base models with and without image enhancement techniques.Evaluation metrics such as average precision, recall, tracking consistency,and computational efficiency are considered.The results demonstrate that image enhancement techniques have a significantpositive impact on multi-object tracking performance across all four models.LSTM, known for capturing temporal dependencies, exhibits improved trackingaccuracy when combined with image enhancement. Vision Transformer, whichutilizes self-attention mechanisms, benefits from enhanced image quality, resultingin superior performance in challenging visual conditions. Siamese Networksand CNN also show enhanced tracking capabilities when integrated with imageenhancement techniques.