In this paper, we investigate the performance of fusion convolutional neural network (CNN) classifier for image-based kinship verification problem. Two fusion configurations were used for the experiments, early fusion CNN classifier and late fusion CNN classifier. The early fusion configuration of the CNN classifier takes combined two face images as input for verification. The advantages of early fusion configuration are no heavy changes in the classifier architecture and only the first layer that have a different filter size. The late fusion configuration of the CNN classifier formed by creating dual CNN network for extracting the deep features of each face image and classify the kinship relationship using two fully-connected layers. The softmax and angular softmax (a-softmax) loss are used for evaluating the network in the training process with fine-tuning strategy. The classifier then evaluated using large-scale FIW (Family in the Wild) kinship verification dataset consists of 1,000 family and 11 different kinship relationship. Experiments using the 5-fold configuration on FIW dataset show that the ensemble of fusion CNN classifier produces comparable performance with several different state-of-the-art methods.