A Novel Feature Fusion with Self-adaptive Weight Method Based on Deep Learning for Image Classification

In recent years, excellent convolutional neural networks (CNN) have been used to solve a variety of visual tasks, and the image classification is an important role for the most visual task. To improve the performance of classification networks, several recent methods have shown the benefit of extracting deeper features by increasing depth of networks and fusing features by linear connection. But deep features only describe the high-level semantic features and loses the shallow features such as edge contour. And features fusion don’t describe the impact factor of features for results. In this paper, we study shallow and deep features fusion and propose a new architectural unit, which we call the “Self-adaptive Weight Fusion” (SFW) method. SFW adaptively recalibrates the features relation by determining impact factors of shallow and deep features, which is used for classify objects. Experimental results demonstrate that SFW can be transplanted to different networks. In particular, we find that this mechanism can produce significant performance improvements for existing state-of-the-art deep architectures with minimal additional computational cost. In addition, results show that shallow and deep features fusion is beneficial to learn more general and robust features for a network.