基於深度學習之易變形商品一維條碼偵測與雙手臂機器人攤平策略

Detection of Barcode on Deformable Objects Based on Deep Learning and Dual-Arm Robot Flattening Strategy

歐彥宏(Yen-Hung Ou)

指導教授 ： 王學誠

國立交通大學/工學院/工學院機器人碩士學位學程/碩士(2020年)

摘要

在進行用於物流行業中之物品揀取系統的開發時，由於1.)商品種類數以萬計以及2.)商品包裝會不定時進行更換，目前主要發展的做法為利用條碼進行商品識別，但是現有條碼偵測方法僅能處理條碼沒有過度扭曲或遮蔽的情況。儘管大部分產品會將條碼印刷在平坦的包裝上，但仍有少部分商品由於使用之材質使得包裝易變形進而導致條碼的扭曲。在商品上的條碼扭曲時目前只能利用人力介入進行處理而沒有相對應之自動化解決方案。為了達成此任務之自動化，需面對的挑戰為:1.)無開源扭曲條碼資料集可使用，2.)以往相關研究之網路架構不一定能沿用，3.)未有利用機械手臂處理條碼之相關研究。為了處理以上挑戰，本論文自行蒐集一組易變形商品之扭曲條碼資料集，並透過自行設計之網路架構進行訓練，同時設計一攤平策略並利用雙手臂機器人進行攤平處理。本論文之主要貢獻為1.)易變形商品之扭曲條碼訓練集，2.)基於深度學習之易變形商品扭曲條碼辨識，3.)利用雙手臂機器人之條碼攤平策略，網路模型與攤平策略在後續實驗中與其他方法相比皆擁有較優良之結果。

關鍵字

雙手臂 ； 條碼偵測 ； 深度學習 ； 易變形商品 ； 語意分割

並列摘要

During the development of the item picking system used in the logistics industry, due to the two major reasons: 1.)tens of thousands of product types, 2.)the packaging of products being changed from time to time, current technique uses barcodes for product identification. However, the existing barcode detection methods can only deal with situations where the barcode is not excessively distorted or obscured. Although most products' barcode is printed on a flat package, there are still a small number of products due to the use of materials that make the packaging easy to deform and cause distortion of the barcode. At the same time, such situations can only be handled by human intervention, and there is no corresponding automated solution. In order to achieve the automation of this task, the challenges to be faced with are: 1.) No open-source twisted barcode data set is available, 2.) The network architecture of previous related studies cannot be determined to be usable, and 3.) There is no relevant research that use robotic arms for barcode processing. In order to deal with the above challenges, I collect a dataset composed of distorted barcode on deformable products, and trains it through a self-designed network architecture, and then uses a dual-arm robot to perform flattening processing by a self-designed strategy. Contributions of this thesis are 1.) Twisted barcode training set for deformable products, 2.) Twisted barcode recognition for deformable products based on deep learning, 3.) Barcode flattening strategy using a dual-arm robot. The network model and flattening strategy is compared with other methods in the later experiments, which all have better results.

並列關鍵字

Dual-arm ； Barcode Detection ； Semantic Segmentation ； Deep Learning ； Deformable Objects

參考文獻

[1] Q. Zhao, F. Ni, Y. Song, Y. Wang, and Z. Tang, “Deep dual pyramid network for barcode

segmentation using barcode-30k database,” arXiv preprint arXiv:1807.11886, 2018.

[2] P. Bodnár and L. G. Nyúl, “Improving barcode detection with combination of simple detectors,”

in 2012 Eighth International Conference on Signal Image Technology and Internet

Based Systems. IEEE, 2012, pp. 300–306.