Enhancing the performance of convolutional neural networks based on preprocessing of datasets

Abstract

Image Recognition is a very challenging task in the various field of computer vision. Convolutional neural networks (CNN) has led to very good performance on a variety of problems in the fields of visual recognition. Although CNNs have achieved great success in experimental evaluations, there are still lots of issues that deserve further investigation. In this research, we proposed a method to improve the performance of a convolutional neural network based on preprocessing of the training and testing sets. We used three different databases Oliva & Torralba, ImageNetDogs and Caltech 256 to train three well-known CNNs AlexNet, GoogleNet and ResNet. Highest performance were obtained to the 70/30 ratio for the training and test set, when the Oliva & Torralba database were used with grid method. Two types of tests were conducted; first test with standardization which limits the all classes of the database to the class that contains the minimum amount of images, and second test with complete database. Results showed that standardizing a database lowers performance. Further, in test 1, it can be seen that the recognition rate for the class with the highest number of samples in Caltech 256, Clutter, was lower and on the contrary, the success rate for classes with fewer samples such as the golden-gate-bridge, harpsichord, scorpion-101, sun ower-10, top-hat were high. Which confirms that the bias towards the Clutter class is diminishing. Test 1 increased the success rate of 106 classes, while decreased for 143 classes. This proved that the best results in terms of performance are obtained when complete databases are used.