This paper introduces a new channel pruning method to accelerate very deep convolutional neural networks (CNNs). The method uses an iterative two-step algorithm based on LASSO regression for channel selection and least squares reconstruction to minimize reconstruction error. It is effective for both single-layer and multi-layer networks, reducing accumulated error and enhancing compatibility with various architectures. The pruned VGG-16 model achieves a 5× speed-up with only a 0.3% increase in error. The method also accelerates modern networks like ResNet and Xception, with 1.4% and 1.0% accuracy loss respectively under 2× speed-up. The approach is implemented at inference time, avoiding the need for retraining and is efficient on both CPU and GPU. The method outperforms previous state-of-the-art approaches in terms of speed and accuracy. The paper also discusses the application of the method to multi-branch networks like ResNet and Xception, and evaluates its performance on various datasets including ImageNet, CIFAR-10, and PASCAL VOC. The results show that the method achieves significant speed-up with minimal accuracy loss, and is more efficient than training-based approaches. The approach is generalizable and can be combined with other techniques like tensor factorization and spatial factorization to further improve performance. The paper concludes that the proposed method is effective for accelerating deep CNNs while maintaining accuracy, and that further research is needed to integrate the method into training time for improved efficiency.This paper introduces a new channel pruning method to accelerate very deep convolutional neural networks (CNNs). The method uses an iterative two-step algorithm based on LASSO regression for channel selection and least squares reconstruction to minimize reconstruction error. It is effective for both single-layer and multi-layer networks, reducing accumulated error and enhancing compatibility with various architectures. The pruned VGG-16 model achieves a 5× speed-up with only a 0.3% increase in error. The method also accelerates modern networks like ResNet and Xception, with 1.4% and 1.0% accuracy loss respectively under 2× speed-up. The approach is implemented at inference time, avoiding the need for retraining and is efficient on both CPU and GPU. The method outperforms previous state-of-the-art approaches in terms of speed and accuracy. The paper also discusses the application of the method to multi-branch networks like ResNet and Xception, and evaluates its performance on various datasets including ImageNet, CIFAR-10, and PASCAL VOC. The results show that the method achieves significant speed-up with minimal accuracy loss, and is more efficient than training-based approaches. The approach is generalizable and can be combined with other techniques like tensor factorization and spatial factorization to further improve performance. The paper concludes that the proposed method is effective for accelerating deep CNNs while maintaining accuracy, and that further research is needed to integrate the method into training time for improved efficiency.