Improving Backbones Performance by Complex Architectures

Recently, Convolution Neural Networks (CNNs) have achieved great success in computer vision. To further boost the performance, the depth of the backbone network is continuously increased, which improves the capacity of feature learning but also brings the heavy burden in computation. To address the issues, this paper introduces a complex convolution method to systematically improve the performance of the backbone network. Our contributions are three-fold: 1) the complex architecture backbone network can improve the classification performance without increasing or even reducing the number of parameters; 2) for the detection task, the complex architecture backbone network can improve the ability of feature map extraction, at the same time our joint bounding box generation method using both real and imaginary parts of complex features can obviously improve the object detection ability. 3) the proposed method has a strong generalization ability for both detection and classification tasks. We have achieved significant performance improvements in both classification and detection tasks, which validate the effectiveness of our methods.