Study of convolutional networks for image classification with applications in aerospace engineering

Abstract

The growing demand for lightweight and efficient aerospace structures has motivated the exploration of Artificial Intelligence (AI) in structural analysis workflows. This thesis investigates the potential of image-based Machine Learning (ML) techniques by manually implementing a simplified, fixed-filter convolutional architecture (pseudo-CNN) in MATLAB, intentionally avoiding high-level ML libraries to gain a deeper understanding of fundamental concepts. The main objective is to compare the performance of a fully-connected neural network (FC-NN) with that of a pseudo-CNN that applies handcrafted filters before classification. Both models are tested on image classification tasks using the MNIST digit dataset and a set of colored animal images. The results aim to highlight the benefits of spatial feature extraction—even without trainable convolution filters—over raw-pixel-based input processing. While this project does not implement a fully trainable CNN, nor does it directly tackle structural optimization tasks, a discussion is provided on how such pre-processing strategies may be conceptually adapted to assist in aerospace applications such as defect detection or structural inspection. The project therefore serves as a pedagogical and exploratory foundation for future work at the intersection of AI and aerospace engineering.