Dahuasecurity.com utiliza cookies y tecnologías similares. Dahua utiliza cookies funcionales para garantizar que sus sitios web funcionen correctamente y cookies analíticas para que su experiencia de usuario sea óptima. Las cookies de terceros también pueden recopilar datos fuera de nuestros sitios web. Al hacer clic en & quot; De acuerdo& quot; o al continuar utilizando este sitio web, usted da su consentimiento para la configuración de cookies y el procesamiento de los datos personales involucrados. Más información sobre nuestra declaración de cookies


Enabling a Safer Society and Smarter Living

Dahua Technology Ranks 1st in Two ReCTS Tasks


Hangzhou, China / April 22, 2021. Dahua Technology, a world leading video-centric smart IoT solution and service provider, has successfully secured the top spots of two major tasks in the recent Robust Reading Competition on Reading Chinese Texts on Signboard (ReCTS), dominating the character recognition and line recognition categories.

Robust Reading is a research area that deals with the detection and recognition of textual information in images under unconstrained settings. Since 2011, Robust Reading Competition has been organized with challenges that cover a wide range of real life scenarios. Each challenge is set up around different tasks. ReCTS is one of these heaps of challenges typically associated with the International Conference on Document Analysis and Recognition (ICDR).

ReCTS dataset includes 25,000 labeled images that are collected through phone cameras under uncontrolled conditions. It mainly focuses on Chinese texts on restaurant signboards. The dataset is split into a training set and a test set. The training set consists of 20,000 images, and the test set consists of 5,000 images. This year, Dahua Technology dominated two major tasks of ReCTS.

ReCTS Task 1: Character Recognition in a Sign Board

(source: rrc.cvc.uab.es / March 24, 2021)

Task 1 involves character recognition in a sign board. It aims to recognize characters from a cropped character image. As baseline, the Dahua team (DH_OCR) used EfficientNet series which were trained with different depth and different width. Synthetic samples generated by Dahua’s own algorithm were also utilized for this task. To balance the data, the samples were processed using smooth, cut and rotate methods. The model itself was trained with ReCTS training data and synthetic data. For this task, the team received a final result of 97.73%, placing them on the top of the list.

ReCTS Task 2: Text Line Recognition in a Sign Board

(source: rrc.cvc.uab.es / March 22, 2021)

Task 2 tackles text line recognition in a sign board. The cropped text line images and the coordinates of the polygon bounding boxes in the images are also given in this particular task. The training data of the Dahua team includes ReCTS, LSVT, RCTW, ART and some high quality artificial synthetic data. CRNN framework were used for text recognition, together with different structures of multi-scale feature extraction backbone such as SAResNET. The team also used multi-model fusion to predict the final result. As a result, the Dahua team ranked 1st on this task, garnering a total score of 96.93%.

Automatic detection and recognition of texts in natural scenes has been receiving increasing attention due to its wide range of applications. It is also a significant requirement for several content-based image analysis tasks. The ReCTS recognitions that Dahua recently received are fruits of the company’s years of continuous innovation in this field and serve as proofs of its dedication to AI technology breakthroughs. With its mission of “Enabling a safer society and smarter living”, Dahua Technology will continue to focus on “Innovation, Quality and Service” to serve its partners and customers around the world.