Application of Factorial Design in Image Processing Using CNN

Convolutional Neural Networks (CNNs) are now dominant deep learning techniques, with phenomenal performance on an extremely diverse set of image processing applications. Their success, however, is radically dependent on proper and accurate adjustment of hyperparameters like the optimizer and learning rate. Even though heuristic-based methods or trial-and-error methods are most likely to produce acceptable results on familiar datasets, it remains a major problem to find the optimal hyperparameter settings for new or unknown image datasets. This information deficiency may cause models to be less than optimal, make computations more expensive, and take longer to deploy on actual systems. This work suggests the use of factorial design, a statistical method of experimental design and analysis that is strict, to systematically and effectively optimize CNN hyperparameters. Factorial design enables formal examination of interactions between leading hyperparameters in a structured way so that optimal settings can be established with lower computational expense than usual. The CIFAR-10 dataset was used as the benchmark for experiments as the validation set for experiments. The experiment compared the effects of three optimizers, Adam, SGD, and RMSprop, with three learning rates, 0.0001, 0.001, and 0.01. Uniform kernel size and padding were employed across configurations to ensure architecture consistency. The CNN architecture used for testing included three blocks of convolutions, which were each preceded by batch normalization, max pool, and dropout layers, and then concluded by a dense fully connected layer. 75 epochs were trained for each of the models, and their performance was tested with respect to validation accuracy and other primary metrics. Among all the combinations that were experimented with, the Adam optimizer with a learning rate of 0.001 performed best, achieving 78.69% validation accuracy. The achievement indicates the prowess of adaptive optimization algorithms in the optimization of model learning. Additional findings were derived using confusion matrix evaluation and visual analysis of predicted outcomes, which unveiled class-wise strengths and weaknesses, exhibiting an inside-out understanding of model behaviour. This study establishes that the adoption of factorial design in hyperparameter tuning of CNNs can markedly simplify experimentation, improve model precision, and streamline computational resources. Future enhancement is likely to happen in the extension of training epochs, using more profound architectures, and leveraging higher-level approaches such as transfer learning, data enrichment, and voting-based methods. In summary, this paper proposes a statistically sound and computationally effective approach to the optimization of CNNs when dealing with unknown or less-surveyed image datasets.