Hybrid optimization strategy of 'Network-in-Network' model for image recognition

In reality, an information processing system needs to receive data from various sensors. This is called a hybrid information system. Image recognition is widely applied to hybrid information systems. Recently, Network in Network (NIN) as a special deep learning model has obtained impressive results on image recognition. However, there is a lot of room for performance improvement, since gradient-based optimization methods may fall into local minima. To get a high recognition rate, it is essential to explore the solution space further. In this paper, we propose a hybrid optimization method for NIN in order to improve the image recognition rate. A fine-tuning strategy was employed to train NIN by using particle swarm optimization followed by a pre-training stage based on the gradient method. Several strategies were introduced in order to avoid the deterioration of the solution. The results showed that the present method can effectively increase the image recognition rate and be further used in a hybrid information system.