Two-Stage Ensemble of Deep Convolutional Neural Networks for Object Recognition

Deep Convolutional Neural Networks (CNNs) have been widely used for object recognition in images. However, due to the complexity in setting the structures and related parameters for deep CNNs, performance of individual CNN models may vary largely and lack robustness. In realizing this problem, this paper proposed a two-stage ensemble of deep CNNs for object recognition. To increase model diversity for ensemble, multiple basic CNN models/structures are first pre-defined. For each basic CNN model, multiple rounds of training are conducted based on sub-sampling the training dataset. In the first stage of ensemble, multiple outputs from each basic CNN model is integrated using the MinMax median. This is followed by the second stage of ensemble to combine the outputs from all basic CNN models. The proposed method was implemented and experiments were carried out on Kaggle's `Statoil/CCORE Iceberg Classifier Challenge' image data for iceberg and ship recognition, as well as on `Northeastern University surface defect dataset' for surface defect classification problem. The experimental results showed that the proposed ensemble method outperformed the individual CNN models, and achieved the-state-of-the-art performance as compared to the best submission to the Kaggle's challenge using CNNs.