FlexReader: Towards Adaptive Decoding Paths for Efficient Arbitrary-Length Scene Text Recognition

Zhibin Ma; Pengwen Dai; Zhigang Yao; Xugong Qin; Wenqi Ren; Xiaochun Cao

cs.CVOct 30, 2026

FlexReader: Towards Adaptive Decoding Paths for Efficient Arbitrary-Length Scene Text Recognition

Authors: Zhibin Ma, Pengwen Dai, Zhigang Yao, Xugong Qin, Wenqi Ren, Xiaochun Cao

Organizations: Shenzhen Campus of Sun Yat-sen University · Shenzhen Key Laboratory of Adversarial Artificial Intelligence · Chongqing University · Nanjing University of Science and Technology

Abstract

Existing Scene Text Recognition (STR) models are inherently coupled with a specific decoding strategy by their training paradigms, limiting their flexibility in deployment across diverse recognition scenarios. In this paper, we propose FlexReader, an STR model that supports multiple decoding strides, enabling a single model to achieve an optimal accuracy-efficiency trade-off across different datasets. Specifically, we first introduce a Sub-string Perception module to perceive the target sub-string from visual features, followed by multiple parallel decoding blocks that independently decode each character in the sub-string. During inference, FlexReader supports decoding with arbitrary strides within the perceived sub-string range. Furthermore, we propose Adaptive Prefix Decoding, which adaptively selects the decoding stride at each step based on the model's confidence in the decoded characters, thereby identifying the optimal decoding path for each input. Experimental results demonstrate that FlexReader exhibits robust recognition capability across a diverse range of benchmarks, including those featuring long text, curved text, and occlusions. It consistently outperforms existing decoding strategies in recognition accuracy while achieving up to 6-flod decoding speedup over autoregressive decoding.

Figures & tables

Figure 1: Left: Comparison of different decoding methods. Right: A representative recognition example, where our method adaptively selects the optimal decoding path.

Figure 2: The overview of the proposed FlexReader . (a) The architecture of the Sub-string Perception module. (b) The implementation of Multi-branch Prediction and the loss computation strategy. Each branch independently computes the loss with its corresponding target, and the losses are then aggregated to obtain the final objective. (c) Visualization of the multi-head cross-attention in Sub-string Perception. (d) The decoding procedures of Fixed-step Decoding (left) and Adaptive Prefix Decoding (right).

$L$	$s$	APD	Union14M-Benchmarks								Time (ms)
$L$	$s$	APD	CUR	MO	ART	CTL	SAL	MTW	GEN	Avg	Time (ms)
4	1		92.7	91.7	80.5	83.0	87.8	87.6	84.3	86.80	58.39
	2		92.5	91.5	79.8	82.3	87.7	86.7	83.9	86.34	34.15
	4		91.8	90.0	80.0	81.8	86.9	86.2	83.5	85.74	20.91
	-	$\checkmark$	92.6	91.1	80.3	82.8	87.5	87.8	84.3	86.63	23.97
8	1		92.5	90.8	78.1	82.1	88.1	86.2	83.8	85.94	61.18

Table 1: Comparison under different decoding configurations.

L

denotes the sub-string perception lengths used during training, while

s

denotes the decoding step size used during inference.

Decoder	Accuracy				Time (ms)
Decoder	Short	Medium	Long	Avg	Short	Medium	Long
ARM	89.7	86.5	72.4	82.9	64.92	109.53	193.23
PDM	87.9	81.7	67.6	79.1	6.30	6.53	7.23
MDM-LC	88.9	83.6	67.6	80.1	16.88	17.59	30.74
FLRE (L=4) †	89.0	85.3	74.3	82.9	17.37	27.29	46.46
FLRE (L=4) ∗	89.6	86.9	73.3	83.3	18.93	29.81	50.62

Table 2: Comparison of different decoding methods. † denotes decoding with the maximum step size, while ∗ denotes decoding with APD. All decoding times are averaged over five repeated measurements on a single A100 GPU.

Figure 3: Left: Text length vs. accuracy for different models. Right: Text length vs. decoding time for different models. All decoding times are averaged over five repeated measurements on a single A100 GPU.

	Accuracy				Time (ms)
	CoB	U14M-B	MCB	Avg	CoB	U14M-B	MCB
$\eta=0$	96.5	87.8	83.4	89.2	7.44	8.39	8.91
$\eta=0.1$	96.5	87.8	83.4	89.2	8.15	8.49	8.97
$\eta=0.3$	96.5	87.8	83.8	89.4	8.89	8.08	9.61
$\eta=0.5$	96.8	88.5	84.8	90.0	8.86	9.17	10.96
$\eta=0.7$	97.0	88.8	85.3	90.4	9.66	10.78	12.29

Table 3: Ablation study of the threshold

\eta

. All decoding times are averaged over five repeated measurements on a single A100 GPU.

Figure 4: Recognition cases of low-confidence character correction by FlexReader. Red denotes incorrect characters.

Figure 5: The visualization results of the cross-attention in the final layer. From top to bottom are the average attention across all attention heads and the attention maps of the 12 individual attention heads, respectively; from left to right are the results at each decoding step.

Method		Common Benchmarks							Union14M-L Benchmark								Params ( $\times 10^{6}$ )
Method		IC13	SVT	IIIT	IC15	SVTP	CUTE	Avg	CUR	MTO	ART	CTL	SAL	MTW	GEN	Avg	Params ( $\times 10^{6}$ )
CTC	CRNN [ 41 ]	91.8	83.8	90.8	71.8	70.4	80.9	81.58	19.4	4.5	34.2	44.0	16.7	35.7	60.4	30.70	8.3
	SVTR † [ 9 ]	96.7	96.7	98.2	88.6	91.3	96.5	94.70	79.3	69.4	67.1	68.8	75.8	76.8	77.0	73.48	18.1
	SVTRv2 [ 11 ]	98.7	98.0	99.2	91.0	93.6	99.0	96.58	91.0	89.5	78.7	81.3	85.9	85.1	82.3	84.91	21.0
PDM	SRN [ 55 ]	97.5	96.3	97.2	87.9	90.9	96.9	94.45	78.1	63.2	66.3	65.3	71.4	58.3	76.5	68.43	51.7
	VisionLAN [ 46 ]	97.1	95.8	98.2	88.6	91.2	96.2	94.50	79.6	71.4	67.9	73.7	76.1	73.9	79.1	74.53	32.9

Table 4: Results on English benchmarks tested against existing models when trained on the real-world Union14M-Filter training set. Yellow and Blue values denote the 1st and 2nd results in each column. † denotes that the result is obtained by training the model on Union14M-Filter using the code they released. ‡ indicates that we use Union14M-Filter and UnionST as the training sets. The data presented in the table corresponds to Word Accuracy Ignore Cases (WAIC).

Scene Text Recognition

Scene Text Recognition reads text in natural images and video, including distorted, multilingual, low-resolution, and stylized writing.

See all 29 papers

Out-of-Length Scene Text Recognition: A Two-Axis Diagnosis and a Training-Free Fix
Jul 25, 2026Zobeir Raisi, John ZelekScene Text Recognition Length Generalization
aElectrical Engineering Department, Chabahar Maritime University, Chabahar, Iran · bDepartment of Systems Design Engineering, University of Waterloo, Waterloo, N2L 3G1, ON, Canada
Can Scene Text Recognition Read Rare Compositions?
Sep 1, 2026Genpei ZhangScene Text Recognition Optical Character Recognition
University of Wisconsin–Madison
2D Rotary Position Embedding for Scene Text Recognition with Transformers
Jul 15, 2026Zobeir RaisiRotary Positional Embeddings Cross-Attention
Electrical Engineering Department, Chabahar Maritime University, Chabahar, Iran.

Jul 25, 2026J/K move · Enter open · S save

Out-of-Length Scene Text Recognition: A Two-Axis Diagnosis and a Training-Free Fix

Zobeir Raisi, John Zelek

aElectrical Engineering Department, Chabahar Maritime University, Chabahar, Iran · bDepartment of Systems Design Engineering, University of Waterloo, Waterloo, N2L 3G1, ON, Canada

Scene Text Recognition Length Generalization Sequence Alignment

Open paper AskPDF arXiv

Scene Text Recognition (STR) models are trained almost exclusively on word crops of at most 25 characters, yet real deployments (signage, product labels, dense captions) require reading much longer text. This paper diagnoses that failure and then closes it. The diagnosis separates out-of-length failure into two simultaneously extrapolating axes (the encoder's width axis and the decoder's time axis) and shows that encoder width, not decoder length, is the dominant failure mode. Representation-side fixes bring only partial relief: training-free rotary rescalings recover at most 2-4 points of character error rate (CER), and a weighted fine-tuning recipe recovers 6-8 points while improving standard-benchmark accuracy, yet word accuracy on the Long Text Benchmark (LTB) stays near zero, because the residual gap lies in the decoding mechanism rather than the representation. We then close that gap at inference time, on an unmodified word-level checkpoint: the long image is sliced into overlapping crops at the model's training width, each decoded independently and in-distribution, and the reads stitched by geometry-anchored edit-distance alignment. This procedure reaches 42.79-43.05% bucket-average word accuracy on LTB across two base checkpoints, matching the published state of the art (41.57%) and beating it by 11-12 points on the hardest bucket, at wall-clock parity with plain decoding; applied unchanged to the public PARSeq checkpoint it reaches 47.11%. Once chunking is applied fine-tuning no longer helps: the decoding-side fix alone matches purpose-built architectures. We release the diagnosis harness and implementation.

Abstract

Scene Text Recognition

Explore similar work

Out-of-Length Scene Text Recognition: A Two-Axis Diagnosis and a Training-Free Fix

Can Scene Text Recognition Read Rare Compositions?

2D Rotary Position Embedding for Scene Text Recognition with Transformers