Design of an Automated CNN Composition Scheme with Lightweight Convolution for Space-Limited Applications

The emergence of the CNN network has enabled many networks for image object recognition, object segmentation, etc., and has brought amazing results to image processing tasks, including MaskRCNN [4] and YOLO [8]. These networks can achieve comparable performance by stacking Convolutional Layers, as layers go deeper, the performance is improved as well. Although deeper convolution layers make the performance of the entire network better, the huge parameters of the networks makes it difficult to implement the network on embedded systems with constrained hardware resources. Therefore, if these networks are to be run on devices with resource constraint, the structure of the network must be lightweight. Usually the most prevalent way to reduce the weight of the network is to modify the network structure, but the design of the network structure has its own philosophy. Any changes to the structure of the network will compromise the performance of the network. We propose a method to automatically substitute the Convolution architecture in the network without changing the network architecture, thereby reducing the parameter of the network while ensuring the performance of the network.