When humans look at a page of text or a crowded street sign, we instantly perceive individual words and lines as distinct entities. For deep learning models based on semantic segmentation, however, this task is deceptively difficult.
In standard semantic segmentation, a network classifies each pixel into a category, such as "text" or "background." While this approach excels at identifying where text exists, it struggles in dense or cluttered environments. When two text lines or words are positioned in close proximity, their predicted pixel masks frequently bleed into one another. This phenomenon is known as instance adhesion or the "sticky text" problem. Once these distinct instances merge into a single connected component, downstream optical character recognition (OCR) engines fail, as they cannot determine where one word ends and the next begins.
To solve this, researchers have shifted from naive pixel classification to instance-aware segmentation strategies. The core intuition behind these solutions is simple: introduce geometric or topological boundaries that force the network to preserve a gap between adjacent instances.
The Core Challenge: Why Proximity Causes Merging
To understand why standard networks fail, consider how a typical Fully Convolutional Network (FCN) learns. During training, the network is penalized based on pixel-level errors. If two text lines are separated by a gap of only one or two pixels, the spatial downsampling inherent to deep convolutional layers (via pooling or strided convolutions) blurs this boundary. The network loses the high-frequency spatial details needed to resolve the gap, causing the probability maps of neighboring instances to merge into a single continuous blob.
As highlighted by Liao et al. in their work on Differentiable Binarization, this issue becomes even more acute in "nested" scenarios—such as a small word printed inside a larger character or logo. Standard segmentation frameworks lack the mechanism to decouple these overlapping supervision regions, treating them as a single semantic mass rather than distinct instances.
To combat instance adhesion, the field has converged on three primary mechanisms: kernel shrinking, explicit border learning, and continuous distance fields.
Mechanism 1: Kernel Shrinking and Differentiable Binarization
One of the earliest and most intuitive workarounds for instance adhesion is to artificially shrink the ground-truth targets during training. Instead of training the network to predict the full bounding polygon of a word, the target is scaled down to a central "kernel." Because the kernels of adjacent words are physically separated by a wide margin, the network can easily predict distinct, non-overlapping regions. During post-processing, these kernels are expanded back to their original shapes using algorithms like polygon offsetting.
However, traditional binarization (turning a probability map into a hard 0-or-1 mask via thresholding) is non-differentiable. This means the step that actually separates the instances cannot backpropagate gradients to help the network learn better features.
Standard Pipeline: [Probability Map] ---> (Hard Thresholding) ---> [Merged Mask] (No Gradients)
DBNet Pipeline: [Probability Map] ---\
---> (Differentiable Step) ---> [Separated Mask] (End-to-End)
[Threshold Map] ---/
To resolve this, Liao et al. proposed Differentiable Binarization (DB). Instead of using a fixed global threshold, the network concurrently predicts a probability map and an adaptive threshold map for every pixel. The binarization is formulated as a modified sigmoid function that approximates a step function but remains fully differentiable. This allows the network to automatically increase the threshold in the narrow gaps between adjacent words, effectively carving out a separation boundary during end-to-end training.
Mechanism 2: Explicit Border Learning and Repulsive Forces
A second approach treats the space between text instances as an explicit entity that the network must actively detect. If the network can learn to identify the "fence" surrounding each word, it can use that fence to block adjacent masks from merging.
Wu et al. introduced this concept in Self-Organized Text Detection with Minimal Post-processing via Border Learning by framing text detection as a three-class problem: text region, background, and text border. By penalizing the network heavily when it misclassifies a border pixel, the model naturally learns to suppress the probability map at the exact boundary where two text lines meet. Similarly, Polzounov et al. utilized a border-aware category in WordFence to prevent adjacent words from bleeding together, eliminating the need for complex, hand-crafted post-processing heuristics.
Taking this intuition a step further, Liu et al. introduced the concept of a Repulsive Text Border in An Accurate Segmentation-Based Scene Text Detector with Context Attention and Repulsive Text Border. Instead of just classifying the border, their framework applies a repulsive loss function that treats adjacent text instances like like-charged magnets. The loss penalizes the network if the predicted boundaries of two neighboring text blocks begin to overlap, actively pushing the instances apart in the feature space.
Mechanism 3: Distance Fields and Topographic Maps
While kernel shrinking and border learning rely on discrete boundaries, a third paradigm transforms the problem into a continuous topological space. Instead of predicting a flat binary mask, the network is trained to predict a continuous distance field or a center-border probability map, where the value of each pixel represents its distance to the nearest text boundary.
In this formulation, a text instance resembles a mountain: the center line of the text is the peak (highest probability), and the values smoothly slope down to zero at the edges.
Cross-section of two adjacent text lines:
Peak (Line 1) Peak (Line 2)
/\ /\
/ \ / \
________/ \____Valley (Gap)____/ \________
Zhu et al. leveraged this topographic intuition in TextMountain. Because each text instance forms its own distinct mountain peak, even if two text lines are physically touching at their bases, they remain separated by a "valley" in the prediction map. A simple hill-climbing or watershed algorithm can then group pixels by tracing them back to their respective local maxima, ensuring clean instance separation.
To make this geometric guidance even more precise, Zheng et al. proposed the Region Multiple Information Perception Network (RMIPN), which forces the network to learn both the distance map and the direction vectors pointing toward the nearest text center. By providing the model with explicit coordinate-aware priors, the system can mathematically decouple instances that are in extreme proximity.
As Wang demonstrated in UHT, initiating the final boundary extraction process from these high-confidence central "seeds" guarantees that the topological separation of nearby text elements is maintained throughout the entire pipeline.
Summary
The evolution of segmentation-based scene text detection is largely a history of finding creative ways to prevent instance adhesion. Whether by dynamically adjusting thresholds via differentiable binarization (Liao et al.), enforcing repulsive boundaries (Liu et al.), or treating text regions as continuous distance mountains (Zhu et al.), the goal remains the same: ensuring that spatial proximity never compromises semantic identity.
Key Foundational Work
A Contour-Based Moving Object Detection (2009) — contributed 2 concept elements to this direction.