Analyzing the effect of hyperparameters in a automobile classifier based on convolutional neural networks

In the recent years the convolutional neural network is used successfully in applications of image classification, due to its deep and hierarchical architecture. The hyper parameters of the convolutional neural networks are of great influence to obtain good results in binary classification without the need of a large number of layers. The activation function, the weights initialization and the sub sampling function are the three main hyper parameters. In the present work 27 models of convolutional neural network are trained and tested with automobile images taken from a surveillance camera. The illumination intensity of the test images are different from the training images, because they were taken from scenes of day, evening and night. We also demonstrate the influence of the mean of the images and the size of the filter kernel. The convolutional neural network model with the best result reached 95.6% of accuracy. The results of experiments show that neural networks predict successfully automobile images with varied illumination intensities overcome the techniques Haar Cascade and the Support Vector Machine.