Text-Pass Filter: An Efficient Scene Text Detector

TL;DR

Text-Pass Filter (TPF) introduces a band-pass-inspired mechanism to segment whole-text regions directly for arbitrary-shaped scene text detection. By learning a per-text pass-feature and a corresponding pass-filter, and augmenting with a Reinforcement Ensemble Unit (REU) for feature-consistency and a Foreground Prior Unit (FPU) for robust foreground discrimination, TPF enables parallel, post-processing-free detection via a filter sieve. The architecture combines a ResNet+FPN backbone, center-point prediction, and a dual-branch feature-filter generator, with a loss function ${\cal L} = \alpha {\cal L}_{fpu} + \beta {\cal L}_{cpt} + \mu {\cal L}_{ffp} + \lambda {\cal L}_{reu}$ that emphasizes both localization and text-feature quality. Evaluations on SynthText, MSRA-TD500, ICDAR2015, Total-Text, and CTW1500 demonstrate competitive or superior accuracy with higher speed than many CNN-based SOTA methods, highlighting strong practical impact for real-time, scalable scene text detection. However, text overlay and potential overdetection from filter strengthening indicate avenues for future integration of higher-level semantic or language cues.

Abstract

To pursue an efficient text assembling process, existing methods detect texts via the shrink-mask expansion strategy. However, the shrinking operation loses the visual features of text margins and confuses the foreground and background difference, which brings intrinsic limitations to recognize text features. We follow this issue and design Text-Pass Filter (TPF) for arbitrary-shaped text detection. It segments the whole text directly, which avoids the intrinsic limitations. It is noteworthy that different from previous whole text region-based methods, TPF can separate adhesive texts naturally without complex decoding or post-processing processes, which makes it possible for real-time text detection. Concretely, we find that the band-pass filter allows through components in a specified band of frequencies, called its passband but blocks components with frequencies above or below this band. It provides a natural idea for extracting whole texts separately. By simulating the band-pass filter, TPF constructs a unique feature-filter pair for each text. In the inference stage, every filter extracts the corresponding matched text by passing its pass-feature and blocking other features. Meanwhile, considering the large aspect ratio problem of ribbon-like texts makes it hard to recognize texts wholly, a Reinforcement Ensemble Unit (REU) is designed to enhance the feature consistency of the same text and to enlarge the filter's recognition field to help recognize whole texts. Furthermore, a Foreground Prior Unit (FPU) is introduced to encourage TPF to discriminate the difference between the foreground and background, which improves the feature-filter pair quality. Experiments demonstrate the effectiveness of REU and FPU while showing the TPF's superiority.

Figures (12)

• Figure 1: Motivation of the designed text-pass filter (TPF). It constructs a unique feature-filter pair for each text, where every feature represents a unique text. Each filter can only allow through the corresponding unique text feature and block the others, which helps detect every text separately in an efficient way.
• Figure 2: Overall architecture of the proposed TPF. It consists of a feature extractor, center point prediction header, feature-filter pair generator, Reinforcement Ensemble Unit (REU), and Foreground Prior Unit (FPU). The extractor includes the backbone and FPN and the corresponding output is a concatenated feature map with the size of $\frac{H}{4}$,$\frac{W}{4}$. The center point prediction header is responsible for locating text instances. REU is designed for encouraging the feature-filter pair generator to produce feature-filter pair with high quality for each text. FPU is introduced to help recognize text center points more accurately. The black flow and gray flow illustrate the forward and backward propagation of the whole training process. Particularly, the gray arrows are the inference-only operators. They bring no extra computational cost for the testing process.
• Figure 3: Structure details of stacked smooth layers and smooth layer. They are composed of 3$\times$3 convolution al layers mainly.
• Figure 4: Structure details of the designed REU. The feature and filter are the outputted feature map and filter map from the feature-filter pair generator. The center point denotes the point coordinates predicted by the center point prediction header. The text features and filters first are sampled from the feature map and filter map according to center point coordinates. The feature consistency and filter recognition ability then are enhanced and strengthened under the guidance of the reinforcement matrix, respectively. In the end, the strengthened filters are combined as a filter sieve by the filter ensemble algorithm (as illustrated in Algorithm \ref{['algorithm1']}) for detecting texts efficiently.
• Figure 5: Visualization of inference process. Structure details of center point prediction header, feature-filter pair generator, REU, and FPU can be found in Fig. \ref{['V2']}, Fig. \ref{['V3']}, and Fig. \ref{['V4']}.