Convolutional Neural Networks (CNNs) have been widely acclaimed for image classification tasks in the last decade. Despite the recent success, training a CNN for large datasets from scratch is still a challenging task due to the long training time and the uncertain iterative learning process of the fully connected layer. In this paper, a novel Convolutional Ensemble Network (CEN) is proposed which can overcome some of the CNN's problems due to slow iterative training process. The proposed network aims to improve two crucial things (1) training time and (2) classification accuracy. In the proposed approach, images are passed to network's convolution and pooling layers to extract the features and then features are passed to an ensemble layer for ensemble weight calculation. In ensemble layer, bootstrapped samples are created randomly using extracted features, further diverse classifiers are trained on each of the sampled features. Finally, the network's output, errors are calculated. The whole process is repeated for different number of iterations and a systematic comparative analysis of results is conducted for standard CNN and proposed CEN. Three benchmark datasets such as MNIST, F-MNNIST and CIFAR-10 are used to evaluate the proposed approach. The accuracy scores achieved for MNIST, F-MNIST and CIFAR10 are 99.08%, 91.5%, 89.76% respectively. Extensive experiments and detailed comparative analysis of results in conjunction with the statistical experiments show that CEN can achieve better accuracy than the standard CNN in lesser iterations.