Tiny Image Classification using Four-Block Convolutional Neural Network

Abstract

The task of classifying images into predefined classes is a major problem in computer vision and artificial intelligence. Deep neural network such as Convolutional Neural Network (CNN) have shown great success in large-scale dataset of high resolution image classification. Here, it is important to note that most real time images may not have high resolution. With the increasing demand of surveillance camera-based applications, the low resolution images are major problem. To overcome this problem, we propose two Four-Block CNN model; one with four-layers and the other one with three-layers. Our proposed Four-block four-layer CNN model contains four convolution layers, first three layers contains 3 × 3 kernel size with stride-1 and fourth layer used with stride-2 for dimensionality reduction.The Four-block three-layer has two convolutional layers with stride-1 and third layer with stride-2. We trained our model on CINIC-10 and CIFAR-10 datasets having low-resolution images. For taking average of whole feature map we are using Global Average Pooling layer as a classifier in both models. To reduce training time complexity, we use non-saturating neurons. Overfitting problem has been addressed by dropout and batch normalization methods. On the validation data, we achieved the best accuracy of 81.62%, 92.21% for CINIC-10 and CIFAR-10 respectively, using Four-block four-layer model. © 2019 IEEE.