Metaheuristics for Deep learning architectures. Application on computer vision problems

DJENAIHI, ELHANI (2025) Metaheuristics for Deep learning architectures. Application on computer vision problems. Doctoral thesis, Faculté des Sciences et de la technologie.

Text PhD_Thesis__14_06_2025.pdf Download (4MB)

Abstract

Deep Neural Networks (DNN) have proven helpful in computer vision tasks, but designing their architecture for optimal performance is challenging. This task demands considerable time and expertise due to the numerous parameter ranges involved. In response to this challenge, our thesis introduces three approaches to address this issue. First, we use a Particle Swarm Optimization (PSO) version known as particle swarm optimization without velocity equation (PSWV). The second method is a hybrid algorithm that combines PSO with Genetic Algorithm (GA), and the third method is MPSO (Mutationenhanced PSO for CNN Architecture Optimization). These techniques minimize human intervention in CNN design, achieve quick convergence, and reduce the time needed to find the best CNN design. All three approaches employ a varying-length encoding strategy to represent particles within the population and integrate new particle updating methods. We compare the proposed algorithms with several contemporary methods in the literature, including 27 recent approaches, and extensively evaluate them on nine benchmark datasets used for classification tasks. Based on empirical findings, the pswvCNN, GAPSO, and MPSO approaches effectively identify CNN structures that perform similarly to standard designs

Actions (login required)

View Item