基於多尺度預測和循環對抗網路的招牌檢測與識別方法之研製;Signboard detection and recognition deep learning modeling based on multiscale prediction and CycleGAN

NCUIR > College of Electrical Engineering & Computer Science > Graduate Institute of Computer Science and Information Engineering > Electronic Thesis & Dissertation > Item 987654321/85057

Please use this identifier to cite or link to this item: http://ir.lib.ncu.edu.tw/handle/987654321/85057

Title:	基於多尺度預測和循環對抗網路的招牌檢測與識別方法之研製;Signboard detection and recognition deep learning modeling based on multiscale prediction and CycleGAN
Authors:	林冠宏;Lin, Kuan-Hung
Contributors:	資訊工程學系
Keywords:	深度學習;物件檢測;招牌辨識;deep learning;object detection;signboard recognition
Date:	2021-01-15
Issue Date:	2021-03-18 17:31:18 (UTC+8)
Publisher:	國立中央大學
Abstract:	物件偵測在電腦視覺任務上是一個很熱門的領域，此技術被使用在許多領域上。為了在提高預測精確度的同時也要保證執行速度在物件檢測上是一個很大的挑戰。有許多專家、學者已致力於這項任務上並提出了許多方法，使得物件檢測的方法日益成熟。在物件檢測中的大多資料集其背景相當複雜，使得模型沒有檢測到目標物件或者發生誤判的情形，為了要解決檢測遺漏有許多方法被提出，例如特徵金字塔網路、多尺度預測和注意力模組等，但極少有方法用以解決將背景誤判為目標物件上。在本文中我們提出了一個兩階段訓練方式的物件檢測模型，用以使用在臺灣街景招牌資料集上，此方法添加了部份語意分割技巧且無須使用到像素間的標記，解決由於大多招牌形狀極為相似而引發的誤判情況。此外我們將此方法進一步的改良使其成為一階段的物件檢測模型，使它的預測結果更加穩定且易於訓練。 ;Object detection is a popular computer vision task in deep learning and the technique is widely used in many fields. To improve the precision of the models while ensuring the inference time is a big challenge. Many experts and scholars have invested in this works and proposed lots of methods to solve this problem, making object detection become more and more mature. The scenes in most object detection datasets are very complicated so that the model cannot detect the objects or it might regard background as an object. To conquer miss detection, lots of methods are proposed like Feature Pyramid Network, multi-scales prediction and attention module. However, there are few methods to prevent the models from misjudging non-objects to objects. In this thesis, we propose a two-phase training method used for Taiwan Street View Signboard Dataset. The model is added with some techniques from segmentation without pixel-to-pixel labeling, solving misjudgments caused by the similar shapes of various signboards. We further improve the method into a one-stage detection model, make the model to be more stable and easier for training.
Appears in Collections:	[Graduate Institute of Computer Science and Information Engineering] Electronic Thesis & Dissertation

Files in This Item:

File	Description	Size	Format
index.html		0Kb	HTML	115	View/Open

社群 sharing

Loading...