Convolutional Neural Networks (CNNs) have achieved much success mainly in areas of computational vision, such as image recognition, classification, object segmentation, and more. The learning process of this type of network generally requires large volumes of data, commonly high-resolution images, and the adjustment of a large number of parameters. The lack of control over the learning process of the model can lead to various problems. One of them is overfitting, which leads the network to a situation where it loses generality, making incorrect forecasts in the presence of new data. Another very common problem is its speed of convergence, which depends on the parameterization of the network: selection of the number of filters per layer, number of convolution layers, and more, where a fine adjustment is very important to avoid excessive computational costs. Understanding the origins of these problems and the ways to prevent them from happening is essential for a successful design. In this paper, we analyze these problems by designing a multiclass classifier among ten categories of images from the Caltech 256 dataset, based on the metrics of accuracy, precision, recall, and loss. To do so, python 3.6, TensorFlow and Keras libraries were used on an RTX 2060 GPU.

No abstract