A Lightweight Convolutional Neural Network using Feature Filtering Module

Most lightweight networks adopt simple feature fusion methods, such as element-wise addition, to achieve real-time speed. However, complex background and noise can affect these methods, making it difficult to locate objects accurately. At the same time, if a large number of channel connections are used to fuse the feature layer, the parameter quantity will increase dramatically. In this work, we propose a new network architecture with dense connection and feature filtering to tackle this problem. We introduce a feature filtering module (FFM) designed to implement feature fusion and filter features adaptively according to the global semantic information so as to reduce feature redundancy and aliasing after feature fusion. We verify the proposed network on CIFAR-10, CIFAR-100, Pascal VOC, and MS COCO. Compared with other state-of-the-art works, experimental results show that our network maintains lightweight while improving system accuracy on object recognition. In addition, our FFM can be applied to other models to improve the performance in object detection.