A novel method for fast arbitrary-oriented scene text detection

Compact and Accurate Scene Text Detector

“…Such compact structure is mainly obtained by adopting Pvanet as a backbone, showing that it is important to carefully choose the backbone. In [14] the ResNet was adopted as a backbone to improve effectiveness without losing inference time. Using ResNet as a backbone led t a better F1 score, but it turned out that it causes efficiency degradation in terms of the number of parameters and FLOPS.…”

“…In this paper, we consider three efficiency factors: inference time, FLOPS, and the number of parameters. Many previous studies using VGGnet or Resnet considered only part of these parameter, for example, only inference time was measured in [10,14]. Such studies are efficient in terms of the inference time, however they are not suitable for mobile environments that may have severely limited resources (e.g., computational power, memory).…”

“…For example, in [10], the decoder simply consists of standard convolutional layers and pooling layers. In [12,14], 1 × 1 convolutions and 3 × 3 convolutions form decoders. The decoder has a huge impact on the performance in terms of the effectiveness and the efficiency.…”

“…In [12], efficient and accurate scene text detection (EAST) was introduced to perform scene text detection using Pvanet [13] as a backbone and standard convolution layers as a decoder. In [14], a similar approach to the above was included, but the backbone was Residual network (Resnet) [15], which is much larger than the Pvanet. In [10] visual geometry group network (VGGnet) was used [16], which is also a large backbone.…”