Enhancing Image Segmentation with Optimized Winograd Algorithm for Convolution Neural Network

---

Abstract

The U-Net architecture has emerged as a popular choice for image segmentation, encompassing convolution, max pooling, upsampling, and concatenation layers. The work aims on enhancing the Winograd algorithm used for convolutions, by introducing a Convolution Pooling Engine that incorporates alterations in the data transformation stage and integrates bias adjustments into the existing adaptations of Winograd convolution. This approach leads to a notable reduction of 7% in sub/add operations. Besides, the convolution computation immediately after an upsampling layer is often inefficient due to redundant data. To address this issue, we propose a novel Upsampling Convolution Engine that results in 25% reduction in add operations. Using HLS4ML flow, we have compared results of custom U-Net IP with HLS4ML IP. These IPs were integrated with MicroBlaze processor on Kintex (KC705 REV 1.2) board. We found that customised IP is having 3.2 times better latency.

---

Keywords

Convolution pooling engine (CPE), upsampling convolution engine (UCE), visual geometric group (VGG)