Hyperparameter search and performance assessment of convolutional neural networks for image classification

As the fields of deep learning, computer vision gain increasing popularity. Image classification is a fundamental task in computer vision that aims to understand images holistically and classify them into specific categories. Various types of image classification algorithms based on convolutional neural networks (CNNs) have been developed and evaluated on multiple image datasets. During the training of many image classification algorithms, several factors can significantly impact the results and even the learning and training process. These factors include the complexity of the image inputs, the optimizer method used, and the parameter tuning techniques applied. In this article, the author conducted experiments using ResNet-18, a residual network learns residual functions to avoid overfitting, with different combinations of learning rates and optimizers. After analyzing the experimental results, it could be observed that the extent of input complexity can indeed affect the accuracy of the models results, as well as its convergence behavior during the training process.