Performance optimization and acceleration of convolutional neural networks in computer vision tasks

Deep convolutional neural networks have made groundbreaking progress in various fields of computer vision. However, the high storage and processing expenses associated with building these networks limit their usefulness in low-powered devices like those found in the Internet of Things and smartphones. Network compression and acceleration techniques have emerged to address this issue by reducing redundancy in the network to compress network parameters and speed up computation. This paper proposes a small-sample convolutional neural network compression and acceleration algorithm. Although pruning algorithms can effectively identify redundancy in the network, they, like most network compression and acceleration algorithms, require a large amount of training data. However, sufficient training data is not always available due to privacy, copyright, and storage issues. To address this, this paper proposes a small-sample convolutional neural network compression and acceleration algorithm based on feature reuse and enhanced knowledge distillation. This algorithm only requires a small number of unlabeled samples to sufficiently compress the network. Specifically, the algorithm first establishes a compressed network prototype by locally reusing the main feature maps of the original network. Then, it further refines the established network prototype features by using an enhanced small amount of samples to detect the input-output structure of the original network from a global perspective. Extensive experiments conducted on benchmark datasets demonstrate the effectiveness of this algorithm.