應用於語意分割的快速雙重注意力拉普拉斯金字塔網絡

Abstract

語義分割是計算機視覺中必不可少但計算成本高的任務。此外，自注意力機制(Self-attention mechanism)可以幫助提取富含的上下文依賴關係的特徵。但是它需要讓類神經網絡額外增加高的計算量。 在這項研究中，我們基於雙重注意力機制(Dual Attention Module)提出了快速雙重注意力機制(Fast Dual Attention Module)，它可以有效率及有效地提取具長距離依賴性關係的訊息。此外，我們也提出了拉普拉斯金字塔解碼器(Laplacian Pyramid Decoder)，它可以有效地從低解析度的語義分割結果還原高頻率的細節特徵並獲得高解析度的語義分割結果。我們將 FDAM 和 LPD 集成到 ESPNet 中，並將我們提出的網絡架構稱為快速雙重注意力拉普拉斯金字塔網絡(Fast Dual Attention Laplacian Pyramid network)。我們在 Cityscapes 數據集上評估 FDALPNet準確率及計算速度。FDA 相對於 DA 在執行時間上降低了 76.68%。LPD 讓 ESPNet 的mIoU score 提升了 5.41%。 FDALPNet 相對於 ESPNet mIoU score 提升了 8.14%。實驗結果顯示FDALPNet 相對於 ESPNet 的準確率有顯注的提升。

Semantic segmentation is an essential yet computationally expensive task in computer vision. Self-attention mechanism can help to capture rich contextual dependencies. However, it requires an even higher computation overhead. In this thesis, we propose a Fast Dual Attention Module (FDAM), which is based on the Dual Attention Module (DAM), that can capture the long-range dependencies information both efficiently and effectively. Besides, we introduce a Laplacian Pyramid Decoder (LPD), which can effectively recover the high-frequency information from a low-resolution segmentation mask. We integrate FDAM and LPD into the ESPNet and call our proposed framework as Fast Dual Attention Laplacian Pyramid network (FDALPNet). We evaluated FDALPNet on the Cityscapes dataset. FDA module is 76.68% less running time than the DA module. LPD improves the mIoU score by 5.41%. The experimental results show that FDALPNet performs favorably against the ESPNet in terms of accuracy. FDALPNet is 8.14% more accurate than ESPNet.