Prof. Milan Tuba

Intelligent Systems for Digital Image Classification

Many modern systems, like self-driving cars, security apps, and automated diagnostics, rely heavily on digital image classification. Over the years, researchers have developed many techniques, but convolutional neural networks (CNNs) have transformed the field, greatly improving classification accuracy. Results of previously advanced research topics are now achievable with CNNS without a lot of effort or time. However, CNNs also introduced new challenges. One big issue is finding the best CNN architecture, which is really tough because there are so many variables to adjust, like the number and type of layers, the number of neurons, kernel size, pooling type, optimization algorithm, padding, and stride. Each model should be fine-tuned for each problem. To handle this, often used method is a trial-and-error approach, or a method called grid search. Since this is an optimization problem, recent studies have also tried metaheuristics like swarm intelligence algorithms. While these methods take a lot of time, they can significantly boost accuracy. This talk will explore the latest advancements and challenges in CNNs, such as tuning hyperparameters.