Memory-Efficient Models for Scene Text Recognition via Neural Architecture Search

Searching a High Performance Feature Extractor for Text Recognition Network

“…• As in NAS algorithms [15], [26], [52], [53], O c contains inverted bottleneck convolution (MBConv) layers [61] with kernel size k ∈ {3, 5} and expansion factor e ∈ {1, 6}. • As in [3], [8], [16], we use an output feature map of size 1×W/4.…”

“…In this experiment, we compare TREFE with the following state-of-the-arts: (i) Bluche et al [80], which uses a deep architecture with multidimensional LSTM to extract features for text recognition; (ii) Sueiras et al [81], which extracts image patches and then decodes characters via a sequence-tosequence architecture with the addition of a convolutional network; (iii) Chowdhury et al [82], which proposes an attention-based sequence-to-sequence network; (iv) Bhunia et al [83], which uses an adversarial feature deformation module that learns to elastically warp the extracted features; (v) Zhang et al [84], which uses a sequence-to-sequence domain adaptation network to handle various handwriting styles; (vi) Fogel et al [85], which generates handwritten text images using a generative adversarial network (GAN); (vii) Wang et al [38], which alleviates the alignment problem in the attention mechanism of sequence-to-sequence text recognition models; (viii) Coquenet et al [13], which replaces the sequential model with lightweight, parallel convolutional networks; and (ix) Yousef et al [33], which does not use a sequential model but instead applies convolutional blocks with a gating mechanism; (x) Shi et al [8], which uses VGG as the spatial model and BiLSTM as the sequential model, and (xi) AutoSTR [37]. We do not compare with STR-NAS [15] (which is concurrent with an earlier conference version [16] of TREFE) as its reported performance is significantly worse. Tables 4 and 5 show results on the IAM and RIMES datasets, respectively.…”

“…Recently, it has been shown that neural architecture search (NAS) [21] can produce good network architectures in tasks such as computer vision (e.g., image classification [22], [23], semantic segmentation [24] and object detection [25]). Inspired by this, rather than relying on experts to design architectures, we propose the use of one-shot NAS [22], [23], [26] to search for a high-performance TR feature [8] fixed vgg [11] BiLSTM -× designed ASTER [3] fixed residual [12] BiLSTM -× GFCN [13] fixed gated-block [13] --× SCRN [14] fixed residual [12] BiLSTM -× STR-NAS [15] fixed searched BiLSTM grad. × NAS AutoSTR [16] two-dim searched BiLSTM grid+grad.…”

“…Concurrently, Hong et al [15] also considered the use of NAS in scene text recognition. However, they only search for the convolution operator, while we search for both the spatial feature and sequential feature extractors with deployment constraints (see Table 1).…”

FAST: Searching for a Faster Arbitrarily-Shaped Text Detector with Minimalist Kernel Representation

“…Despite these developments, these NAS-based models are mainly limited to a few tasks such as image classification and general object detection, leading to the weak generalization ability in other tasks. To compensate for these drawbacks, many researchers explore applying NAS methods to their specific domains, including semantic segmentation (Liu et al 2019a), pose estimation (Xu et al 2021), and scene text recognition (Zhang et al 2020a;Hong, Kim, and Choi 2020). However, there is still rare to extend NAS approaches to text detection.…”