國立台灣科技大學-專利商品網-專利技術-基於密集多尺度網路模型應用於瑕疵檢測

●

技術名稱 Technology

●

發明人 Inventor

郭景明, 張立穎,

●

所有權人 Asignee

國立臺灣科技大學

●

專利國家 Country	申請號 Application No.	專利號 Patent No.	中心案號 Serial No.
中華民國	110115669	I776489	1090050TW0
美國	17/344,911	US 11,657,279B2	1090050US1

點閱數：670

技術摘要：

Nowadays, document segmentation is one of the key problems
in the eld of semantic segmentation. Although this demand applying
deep learning approaches has undergone continuing advancement,
convolutional neural networks are commonly limited at a xed resource
budget. To tackle this problem, we present the high-eciency encoderdecoder
model for pixel-wise document segmentation, namely EDSNet.
The architecture of the encoder network is based on EcientNet. The
decoder stage is composed of a set of convolutions and upsampling layers
to full input resolution feature maps for pixel-wise classication. To
overcome the contour of document features and multi-level feature capability,
this paper proposes an edge supervision network and Densely
Joint Pyramid Module (DJPM) to improve the performance. As a result,
the proposed document page segmentation method achieves 91.22% using
IoU evaluation on the dataset of RDCL 2017. Moreover, a large-scale
dataset for document segmentation, called PPSD2020, was released with
this paper for open access.

相關圖片：