• PDF / 1,082,861 Bytes
• 14 Pages / 439.37 x 666.142 pts Page_size
• 13 Downloads / 226 Views

DOWNLOAD

REPORT

---

Abstract. We propose a scene text detection system which can maintain a high recall while achieving a fair precision. In our method, no character candidate is eliminated based on character-level features. A weighted directed graph is constructed and the minimum average cost path algorithm is adopted to extract line candidates. After assigning three line-level probability values to each line, the final decisions are made according to the line candidate clustering of the current image. The proposed system has been evaluated on the ICDAR 2013 dataset. Compared with other published methods, it has achieved better performances. Keywords: Line-level features · Weighted directed graph model imum average cost path algorithm · Line candidate clustering

1

· Min-

Introduction

Text detection and recognition in natural scenes is a key technology to potential applications such as autonomous navigation, multilingual translation, and assistance to the visually impaired people. While there have been significant advances in recent years, text detection still remains challenging due to the diversity of scene text and the complexity of background. For comprehensive surveys, refer to [8,20,24]. A typical text detection system usually consists of four components, namely, character candidate extraction, false candidate elimination, text line generation, and text line verification. Existing methods for scene text detection can be largely categorized into three groups: methods using the sliding window [7,18,23], methods using the connected components [2,5,15,21], and hybrid methods [6]. Zhang et al. [23] adopted the sliding window which is based on the symmetry property of character groups. In the work [7] text saliency maps were computed by evaluating the character/background convolutional neural network classifier in a sliding window. Tian et al. [18] detected character candidates by combining the sliding window scheme with a fast cascade boosting algorithm that exhibited a high performance. Compared with the methods using the sliding window, methods based on connected components have become popular during the last ten years, since c Springer International Publishing Switzerland 2016  G. Hua and H. J´ egou (Eds.): ECCV 2016 Workshops, Part I, LNCS 9913, pp. 364–377, 2016. DOI: 10.1007/978-3-319-46604-0 27

Scene Text Detection with Adaptive Line Clustering

365

these methods are usually more efficient and relatively insensitive to variations in scale, orientation, font, and language type. SWT [4] and MSER [11] are two representative component-based methods for scene text detection, which constitute the basis of a lot of subsequent works [2,5,15,21]. The method of Neumann and Matas [12] exploited the MSER detector and then classified the detected regions as either a character, a multi-character or the background. In [3] a fast stroke detector was proposed based on an efficient pixel intensity comparison to surrounding pixels. However, the character candidate extraction brings in too many non-text components especially in component-based met