This paper focuses on accelerating the evaluation of convolutional neural networks (CNNs) by exploiting cross-channel and filter redundancy. The authors propose two schemes to speed up convolutional layers, which are architecture-agnostic and can be applied to existing CPU and GPU frameworks. The first scheme approximates the filter set using a linear combination of a smaller basis set of rank-1 filters, while the second scheme factors the convolutional layer into two regular convolutional layers with rectangular filters. Both schemes are optimized using filter reconstruction or data reconstruction methods to minimize reconstruction errors. The proposed methods achieve significant speedups with minimal accuracy loss, demonstrating a 4.5× speedup with a 1% drop in accuracy for a scene text character recognition task, still achieving state-of-the-art performance. The paper also compares the proposed methods to other acceleration techniques and shows their effectiveness in practical applications.This paper focuses on accelerating the evaluation of convolutional neural networks (CNNs) by exploiting cross-channel and filter redundancy. The authors propose two schemes to speed up convolutional layers, which are architecture-agnostic and can be applied to existing CPU and GPU frameworks. The first scheme approximates the filter set using a linear combination of a smaller basis set of rank-1 filters, while the second scheme factors the convolutional layer into two regular convolutional layers with rectangular filters. Both schemes are optimized using filter reconstruction or data reconstruction methods to minimize reconstruction errors. The proposed methods achieve significant speedups with minimal accuracy loss, demonstrating a 4.5× speedup with a 1% drop in accuracy for a scene text character recognition task, still achieving state-of-the-art performance. The paper also compares the proposed methods to other acceleration techniques and shows their effectiveness in practical applications.