Optimization of convolutional neural network parameters for image classification

Convolutional Neural Networks (CNNs) have been widely applied in image classification tasks. CNNs have a large number of parameters and they can produce different classification accuracy for same tasks based on diverse parameters including input window size, filter size, number of layers and number of neurons. The impact of these parameters on CNN accuracy in image classification tasks is investigated and analyzed in this study. A new methodology incorporating CNN for systematically conducting experiments to find the impact of diverse parameters is presented. Two datasets such as benchmark CIFAR-10 dataset and road-side vegetation dataset for real-world applications were selected to conduct this study. The experiments were conducted by varying different network parameters and recording the accuracy. Experimental analysis has shown that changing the number of layers and input window size has significant impact on classification accuracy of CIFAR-10, whereas for roadside vegetation dataset input window size and filter size have maximum impact on classification accuracy. The proposed optimization approach achieved higher accuracy (81%) than the accuracy obtained by Alexnet (77.75%) and PSO-CNN (80.15%) on CIFAR-10 dataset. © 2017 IEEE.