Performance Evaluation of Semantic Segmentation Using Different Encoder–Decoder Architectures

不同編解碼器架構下語義分割之性能評估

Semantic segmentation, also known as dense prediction, is a task in computer vision that is critical for scene understanding and commonly used for pixel-wise labeling of whole images. This paper proposes a new semantic segmentation network architecture dedicated to decoder structural design. Moreover, different mapping mechanisms are introduced as a part of the decoder network. The proposed architecture was tested on the 21 classes of the Pascal Visual Object Classes Challenge 2012 data set. The proposed method has a higher mean intersection over union score and pixel accuracy than those of the U-SegNet, UNet, and ENet models but similar results to those of the SegNet model. Additionally, the effect of using magnification methods in the decoder network on object segmentation performance was investigated.

在計算機視覺中，語義分割是最重要的任務之一，因其對於場景理解至關重要。通常，語 義分割用於整個影像之像素標記，故又稱為密集預測。本文之目的在於提出一種新的語義分割 網路架構，特別著重在解碼器結構的細部設計。本文創新之處在於引入不同的映射機制作為解 碼器網路的一部分。本文所提方法在 21 個類別的 Pascal VOC 2012 資料集上進行測試，其結果 同時與 U-SegNet，UNet 和 ENet 等模型進行比較，實驗結果顯示，本文所提方法具有更高的 mIoU 分數與像素精度，但與 SegNet 相比，其分割效能相當。此外，本文亦探究解碼器網路中 特徵圖放大方法對物體分割性能之影響。