Controllable adaptive pruning for DCNN filters based on significant convolution kernel of input channels

With the development of the Internet of Things (IoT) and autonomous driving, many edge devices need to deploy deep network models to improve their performance. However, deep neural network models typically involve a substantial number of parameters and computational demands. Filter pruning can substantially compress and accelerate the deep network models, transforming them into compact and lightweight versions, and facilitating more efficient deployment on edge devices. In this paper, a controlled adaptive pruning method based on significant convolutional kernels of input channels called CAPSCK is proposed for DCNN filters. First, the significance of each group of input channel convolutional kernels is evaluated. The significant convolutional kernels (SCK) are used to measure filter importance, with interference caused by insignificant convolutional kernels minimized during evaluation. Second, a controllable adaptive pruning (CAP) evaluation is constructed. This assesses pruning sensitivity by utilizing the baseline pruning rate for each layer, which is determined based on filter importance. It reduces the subjectivity caused by manually setting the pruning rate. Experiments of pruning on multiple datasets and different networks show that CAPSCK can effectively compress and accelerate network models. For example, on CIFAR-10 with VGG16, CAPSCK achieves a compression ratio of 12.58× and an acceleration ratio of 4.14×, with only 0.16% drop in accuracy. The compression and acceleration performance on various datasets and networks surpasses several state-of-the-art methods.