Cross Connected Network For Efficient Image Recognition

In this work, we describe a novel and highly efficient convolutional neural network for image recognition, which we term the "Cross Connected Network" (CrossNet). We have creatively introduced the Pod structure, where the feature map of depthwise convolutions can be reused within the same Pod by cross connection. Such a design can make the CrossNet have high performance while less computing resource, especially suitable for mobile devices with very limited computing power. Additionally, we find that depthwise convolutions with large receptive field has better accuracy/computation trade-offs, and further improves CrossNet performance. Our experiments on ImageNet classification and MSCOCO object detection demonstrate that CrossNet can improve the state-of-the-art performance of lightweight networks (such as MobileNets-V1/-V2, ShuffleNets and CondenseNet). We have tested the actual inference time on an ARM-based mobile device. The CrossNet still gets the best performance.