<![CDATA[Convolutional Neural Networks (CNNs) have achieved significant success in the field of computer vision; however, their high computational complexity poses challenges for deployment in real-time applications. This study explores the application of Winograd-based convolution algorithms, specifically F (2,3) and F (4,3), as a means to accelerate CNN inference. Using the VGG-16 architecture as a benchmark, we evaluate the performance of these algorithms in terms of execution time and computational accuracy. Experimental results demonstrate that Winograd F (2,3) reduces runtime by an average of 59.62%, while Winograd F (4,3) achieves a 39.81% reduction compared to standard convolution. Accuracy is assessed using single-precision 32-bit floating-point arithmetic, with results showing that Winograd F (2,3) achieves the lowest maximum element error in six out of nine convolutional layers. These findings indicate that Winograd-based methods offer an efficient alternative to conventional CNN computations, particularly in performance-constrained environments.]]>
Copyrights © 2025
