Character Recognition in Natural Images

By Teofilo de Campos, Bodla Rakesh Babu, Manik Varma. 2009

621 citations recorded by the source corpus.

This paper tackles the problem of recognizing characters in images of natural scenes. In particular, we focus on recognizing characters in situations that would traditionally not be handled well by OCR techniques. We present an annotated database of images containing English and Kannada characters. The database comprises of images of street scenes taken in Bangalore, India using a standard camera. The problem is addressed in an object cateogorization framework based on a bag-of-visual-words representation. We assess the performance of various features based on nearest neighbour and SVM classification. It is demonstrated that the performance of the proposed method, using as few as 15 training images, can be far superior to that of commercial OCR systems. Furthermore, the method can benefit from synthetically generated training data obviating the need for expensive data collection and annotation.

Subfields: Computer Vision and Pattern Recognition, Media Technology

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Computer Vision and Pattern Recognition

In Media Technology

Citations in the corpus

Speeding up Convolutional Neural Networks with Low Rank Expansions (2014, background): “The training dataset consists of 163,222 collected character samples from a number of scene text and synthesized character datasets [1, 2, 5, 17, 21, 34, 42].”
Learning Deep Structured Models (2014, methodology): “raphical model has ﬁve nodes (one per character). We then generated writing variations of each word as follows. We took the lower case characters from the Latin alphabet given in the Chars74K dataset [2], and inserted them in random background image patches by alpha matting (i.e., characters have tr”
Synthetic Data and Artificial Neural Networks for Natural Scene Text Recognition (2014, background): “Here we follow the success of some synthetic character datasets [4, 29] and create a synthetic word data generator, capable of emulating the distribution of scene text images.”
First Session Adaptation: A Strong Replay-Free Baseline for Class-Incremental Learning (2023, methodology): “4.2, we use the 19 datasets of VTAB [42], a low-shot transfer learning benchmark, plus three additional datasets, FGVC-Aircraft [17], Stanford Cars [12], and Letters [6].”
Memory Aware Synapses: Learning what (not) to forget (2017, background): “…[19] for ﬁne-grained aircraft classiﬁcation; VOC Actions , the human action classiﬁcation subset of the VOC challenge 2012 [5]; Letters , the Chars74K dataset [3] for character recognition in natural images; and the Google Street View House Numbers SVHN dataset [22] for digit recognition.”
Selfless Sequential Learning (2018, positive): “The 8 tasks sequence is composed of: Oxford Flowers [31], MIT Scenes [32], CaltechUCSD Birds [43], Stanford Cars [17]; FGVC-Aircraft [29]; VOC Actions [8]; Letters [7]; and SVHN [30] datasets.”
Convolutional neural networks applied to house numbers digit classification (2012, neutral): “Previous approaches in classifying characters and digits from natural images used multiple hand-crafted features [3] and template-matching [15].”
Reading Text in the Wild with Convolutional Neural Networks (2014, methodology): “datasets [9, 57], we create a synthetic word data generator, capable of emulating the distribution of scene text images.”
End-to-End Text Recognition with Hybrid HMM Maxout Models (2013, background): “…is a classiﬁcation problem that is generally addressed with the use of strong clas-siﬁers such as Convolutional Neural Nets (CNNs) (Wang et al., 2011; 2012), deformable parts models (Shi et al., 2013) or manually-engineered feature-extraction followed by a classiﬁer (de Campos et al., 2009).”
A Method for Text Localization and Recognition in Real-World Images (2010, methodology): “On the de Campos et al. Char74k dataset [ 12 ], a highly significant increase in recognition rate from 53% [12] to 72% was achieved.”
Image Retrieval Using Textual Cues (2013, methodology): “It is to be noted that [8] only uses 15 training samples per class.”
SpArSe: Sparse Architecture Search for CNNs on Resource-Constrained Microcontrollers (2019, background): “4 Results We report results on a variety of datasets: MNIST (55e3, 5e3, 10e3) [43], CIFAR10 (45e3, 5e3, 10e3) [39], CUReT (3704, 500, 1408) [60], and Chars4k (3897, 500, 1886) [25], corresponding to classification problems with 10, 10, 61, and 62 classes, respectively, with the training/validation/t”
Deep Neural Decision Forests (2015, neutral): “G50c [33] Letter [10] USPS [14] MNIST [20] Char74k[8] Tree input features 10 (random) 8 (random) 10x10 patches 15x15 patches 10 (random) Depth 5 10 10 10 12 Number of trees 50 70 100 80 200 Batch size 25 500 250 1000 1000 Table 1 .”
End-to-end scene text recognition (2011, methodology): “The work of [6] addressed the cropped character classification problem (1) and showed the relative effectiveness of using generic object recognition methods versus off-the-shelf OCR.”
Alternating Decision Forests (2013, methodology): “We use the G50c data set from [8], the Letter and MNIST data sets from [11], the USPS data set from [18], and the Char74k data set from [6].”
Learning Models with Uniform Performance via Distributionally Robust Optimization (2018, methodology): “We construct our training set as a mixture of MNIST hand-written digits [33] (majority population) and type-written digits consisting of different fonts [31] (minority population).”
Text Detection and Character Recognition in Scene Images with Unsupervised Feature Learning (2011, background): “Common features include edge features, texture descriptors, and shape contexts [1].”
ARDIS: a Swedish historical handwritten digit dataset (2019, background): “The dataset is publicly available for researchers [66, 67].”