Robust water meter reading identification method in challenging scene based on YOLOv11

Abstract

The invention discloses a robust water meter reading identification method under a challenging scene based on YOLOv11, provides a comprehensive solution, and improves the accuracy and robustness of water meter reading identification under the challenging condition by solving the problems of environmental factors and mechanical digital jump. Firstly, a diversified water meter image data set is planned, robust learning of challenging samples is ensured, the adaptability of the model is enhanced, and a solid foundation is laid for model training and generalization. And secondly, the CSFE and the PSFM are integrated in the water meter reading identification model, a YOLOv11n framework is optimized, and the identification robustness in a challenging scene is remarkably improved. Finally, two new indexes TTMRrate and WADRrate are provided for comprehensive and practical performance evaluation, and more accurate identification performance evaluation of the water meter reading identification model is provided. According to the method, 98.3% of mAP at 50 and 96.19% of TTMRrate are realized, and the performance of an existing model based on YOLO on single character and sequence level identification tasks is exceeded.

Description

Technical Field

[0001] The present invention relates to the technical field of water meter reading recognition, and specifically to a robust water meter reading recognition method based on YOLOv11 in challenging scenarios. Background Art

[0002] Water meters are one of the main instruments for measuring water supply in various industries, including civil industries such as residential buildings, hotels, and restaurants, as well as a wide range of industrial fields. Among them, mechanical water meters are roughly divided into two types: one is the pointer type water meter with multiple dials, which requires reading the readings of the pointers on multiple dials and multiplying by the range of each dial to obtain the reading result; the other is the word wheel type water meter with a digital display method, that is, the word wheel type water meter, and the current reading of the water meter is obtained by reading the numbers in the word wheel area.

[0003] Accurate water meter readings are crucial for water resource management, consumption monitoring, and resource optimization. Although traditional manual meter reading methods are still widely used, these methods are error-prone and inefficient, making them less feasible for large-scale deployment. Although smart water meters have automated the process, they also bring challenges such as high costs, frequent maintenance, and dependence on network infrastructure. Image-based recognition provides a promising alternative by combining a camera module with a mechanical water meter to achieve automated meter reading. However, despite the potential of image-based water meter reading, it still faces significant challenges in dealing with environmental changes and mechanical digital jumps, which seriously affect the accuracy of recognition.

[0004] As Figure 1 shown, existing image-based automatic water meter reading still faces two major challenges: environmental factors and mechanical digital jumps. (1) Environmental factors significantly affect the accuracy of image-based water meter reading. Outdoor water meters are often affected by changing light conditions, resulting in severe underexposure or overexposure, which can damage the visibility of the numbers. Underexposure makes the digital edges blurred, while overexposure erases key details, both of which significantly hinder the accuracy of recognition. In addition, glare from protective glass or plastic covers may also obscure the character area, further complicating the recognition process. Over time, stains and wear will accumulate on the dial. Stains will cover the numbers, and wear will distort the integrity of the characters. Optical distortion and segmentation ambiguity introduced by water droplets further increase the difficulty of recognition. (2) In addition to environmental factors, mechanical digital jumps also pose unique challenges. Due to the gear drive mechanism, the digital wheels of word wheel type water meters may display the characteristics of two consecutive numbers during jumps, forming half characters (such as Figure 1b) in. Different from integer characters consisting of a single digit, half characters contain partial features from two adjacent digits, making their classification more challenging. In multi-digit water meter readings, half characters mainly appear on the least significant digit or consecutively in the last few digits. However, their occurrence frequency is much lower than that of integer characters, resulting in a severe class imbalance that limits the model's ability to effectively learn the digit jump state.

[0005] Automatic water meter reading methods can be classified into two categories: sequence-based recognition and character-based recognition. Sequence-based methods utilize convolutional neural networks (CNNs) and sequence learning techniques to directly predict the complete digit reading, eliminating the need for character segmentation. Gomez et al. proposed an end-to-end reading extraction framework based on CNNs, while Yang et al. introduced a fully convolutional sequence recognition network (FCSRN) with an enhanced loss function to enhance sequence modeling. Xiu et al. developed a hybrid model (HRC-mCNNs) that combines regression and classification, which integrates digit consistency constraints, and Zhang et al. combined keypoint localization and recurrent neural networks (RNNs) to directly generate the reading sequence. Zhao et al. further integrated the detection, correction, and recognition modules into a unified pipeline for structured sequence modeling. Although these methods directly generate digit readings, they involve high computational costs and have limited robustness in challenging environments such as extreme lighting conditions, occlusion, and other environmental interferences. In contrast, early character-level recognition methods typically require image preprocessing to reduce noise and distortion, followed by region segmentation using traditional techniques such as edge detection and morphological operations, and finally character classification using template matching or machine learning classifiers. Although these methods improve the efficiency of manual meter reading, they are very sensitive to segmentation errors and environmental changes, resulting in poor performance under complex conditions. To address these challenges, object detection-based methods are increasingly being applied to water meter reading. These methods integrate character localization and recognition in a single model, eliminating the need for an explicit segmentation step. By directly detecting and classifying characters, they reduce segmentation errors while improving recognition accuracy. In the latest advancements in object detection, especially the use of the YOLO series, the performance in handling complex conditions and segmentation challenges has been significantly improved. However, existing methods have not fully solved these challenges because problems related to environmental factors and mechanical digit jumps still exist and further architectural improvements are needed. Summary of the Invention

[0006] (1) Technical Problems to be Solved

[0007] Aiming at the deficiencies of the prior art, the present invention provides a robust water meter reading recognition method based on YOLOv11 for challenging scenarios, which can solve the robustness problem of water meter reading recognition in challenging scenarios.

[0008] (2) Technical Solution

[0009] To solve the above technical problems, the present invention provides the following technical solution: A robust water meter reading recognition method for challenging scenarios based on YOLOv11, comprising the following steps:

[0010] S1. Construct a water meter image dataset, wherein the water meter image dataset includes water meter images of various different challenging scenarios;

[0011] S2. Use the water meter image dataset to train a water meter reading recognition model, wherein the water meter reading recognition model is an improvement based on YOLOv11;

[0012] S3. Use the trained water meter reading recognition model to recognize the water meter reading.

[0013] Preferably, the challenging scenarios include overexposure, underexposure, glare, stains, surface wear, and water droplets.

[0014] Preferably, the water meter image dataset further includes standard water meter images.

[0015] Preferably, step S1 includes the following sub-steps:

[0016] S11. Collect multiple original water meter images;

[0017] S12. Select multiple water meter images of different challenging scenarios from the multiple original water meter images as training samples;

[0018] S13. Use the training samples to train a lightweight classification network;

[0019] S14. Use the lightweight classification network to conduct rough screening from the multiple original water meter images to obtain a roughly screened water meter image dataset;

[0020] S15. Eliminate redundant images in the roughly screened water meter image dataset;

[0021] S16. Construct the water meter image dataset through manual re-screening.

[0022] Preferably, before step S2, it further includes: Marking the water meter images in the training set and validation set of the water meter image dataset with wheel numbers, wherein the wheel numbers include 20 different categories: 10 integer number categories and 10 semi-character categories representing intermediate transition states.

[0023] Preferably, the water meter reading recognition model includes a Challenging Scenario Feature Extractor (CSFE), wherein CSFE includes a local feature extraction component, a Convolutional Additive Token Mixer (CATM), and a Convolutional GLU module.

[0024] Preferably, the water meter reading recognition model includes a depth semantic fusion module PSFM, where PSFM includes feature refinement, cross-frequency attention calculation, and feature fusion.

[0025] Preferably, the robust water meter reading recognition method in challenging scenarios based on YOLOv11 further includes: introducing a meter reading rate that tolerates small-range digital jumps to evaluate the performance of the water meter reading recognition model.

[0026] Preferably, the robust water meter reading recognition method in challenging scenarios based on YOLOv11 further includes: introducing a weight-aware digital recognition rate to evaluate the performance of the water meter reading recognition model.

[0027] (III) Beneficial Effects

[0028] Compared with the prior art, the present invention provides a robust water meter reading recognition method in challenging scenarios based on YOLOv11, which has the following beneficial effects: The water meter image dataset of the present invention includes water meter images of various different challenging scenarios. Using this water meter image dataset to train the water meter reading recognition model can better ensure robust learning of challenging samples and enhance the adaptability of the model. Moreover, the water meter reading recognition model of the present invention improves feature extraction and multi-scale fusion on the basis of YOLOv11. Through the above methods, the present invention can improve the recognition accuracy and robustness of water meter readings in challenging scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 Water meter images corresponding to various challenges faced by automatic meter reading based on images;

[0030] Figure 2 Flowchart of the steps of the robust water meter reading recognition method in challenging scenarios based on YOLOv11 of the present invention;

[0031] Figure 3 Sub-flowchart of the steps of constructing the water meter image dataset of the present invention;

[0032] Figure 4 Overall framework diagram of the water meter reading recognition model of the present invention;

[0033] Figure 5 Architecture diagram of CSFE of the present invention;

[0034] Figure 6 Architecture diagram of PSFM of the present invention;

[0035] Figure 7 Example diagram of the detection performance of the present invention and YOLOv11 n in challenging scenarios;

[0036] Figure 8 This is an example diagram of enhanced character recognition of the present invention and YOLOv11 n in challenging cases. Detailed implementation manners

[0037] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0038] The present invention provides a robust water meter reading recognition method in challenging scenarios based on YOLOv11, including the following steps:

[0039] S1. Construct a water meter image dataset, where the water meter image dataset includes water meter images of various different challenging scenarios.

[0040] Specifically, the above-mentioned challenging scenarios may include overexposure, underexposure, glare, stains, surface wear, and water droplets: Outdoor water meters are often affected by changing lighting conditions, resulting in severe underexposure (underexposure) or overexposure (overexposure), which can damage the visibility of the numbers; underexposure makes the digital edges blurred, while overexposure erases key details, both of which significantly hinder the recognition accuracy. Glare from protective glass or plastic covers may also obscure the character area, further complicating the recognition process. Over time, stains and wear will accumulate on the dial, stains will cover the numbers, and wear will distort the integrity of the characters. In addition, the optical distortion and segmentation ambiguity introduced by water droplets further increase the difficulty of recognition.

[0041] The present invention constructs a dataset containing challenging environmental conditions to enhance the generalization ability and robustness of the water meter reading recognition model. Many existing datasets fail to fully cover extreme lighting changes, surface contamination, and mechanical digital jumps, which limits the adaptability of the model. To solve this problem, the present invention designs a systematic image selection and filtering process to ensure that the above-mentioned water meter image dataset represents difficult cases in a balanced manner while maintaining sample diversity. Preferably, step S1 includes the following sub-steps:

[0042] S11. Collect multiple original water meter images.

[0043] The above-mentioned water meter image dataset is constructed through a multi-stage selection and filtering process; first, collect multiple original water meter images, for example, collect 100,000 original water meter images.

[0044] S12. Select multiple water meter images of different challenging scenarios from multiple original water meter images. For example, select 500 representative water meter images from 100,000 original water meter images to cover various key challenging scenarios, such as overexposure, underexposure, glare, stains, surface wear, and water droplets.

[0045] S13. Use the training samples selected in S12 above to train a lightweight classification network to assist in screening challenging cases. The above lightweight classification network can specifically adopt existing ones such as EfficientNet-B3.

[0046] S14. Coarsely screen from multiple original water meter images using the lightweight classification network to obtain a coarsely screened water meter image dataset.

[0047] S15. Eliminate redundant images in the coarsely screened water meter image dataset in S14 above, while retaining the changes in acquisition conditions. Specifically, time window strategy and K-means clustering can be applied to eliminate redundant images, and there is no excessive limitation here.

[0048] S16. Finally, manually rescreen the dataset after eliminating redundant images in S15 above to construct a water meter image dataset. Through manual rescreening, the balance and high quality of the dataset are ensured, and other key scenarios are supplemented.

[0049] The water meter images of various different challenging scenarios in the water meter image dataset constructed in step S1 can adopt the following distribution: overexposure 12.2%, underexposure 16.6%, glare 19.6%, stains 18.6%, surface wear 12.4%, water droplets 20.6%.

[0050] In addition, in addition to environmental changes, the present invention also introduces classification uncertainty for the digital jump in mechanical water meters to make the water meter reading recognition of the present invention more challenging. Therefore, before step S2, the present invention further includes: performing wheel digit marking (i.e., annotation) on the water meter images in the training set and validation set of the water meter image dataset. Among them, the wheel digits include 20 different categories: 10 integer digit categories (0-9) and 10 semi-character categories (X.5) representing intermediate transition states; among them, for the semi-character categories, when the dial transitions from X to X+1, it is marked as X.5. For example, in the special case when the wheel digit flips from 9 to 0, it is marked as the semi-character category of 9.5 to represent the flip transition. The above wheel digit marking strategy enables the water meter reading recognition model of the present invention to correctly interpret the transition state and reduces the uncertainty of digital interpretation. Preferably, all the above wheel digit markings follow the YOLO format to ensure the consistency and efficiency of model training.

[0051] In addition, preferably, in order to improve the robustness of the model, targeted data augmentation can be performed on multiple randomly selected (e.g., 1,605) water meter images, introducing illumination changes, noise distortion, and geometric changes; specifically, brightness adjustment simulates diverse lighting conditions, fog and speckle noise introduce environmental interference, rotation and perspective distortion consider different camera angles, contrast and saturation adjustment enhance feature discrimination, and image sharpening improves the digital clarity under harsh conditions; these data augmentations expand the diversity of the water meter image dataset and improve the generalization ability of the model to real-world recognition challenges.

[0052] Furthermore, preferably, in addition to including challenging cases (i.e., water meter images of challenging scenarios), the above-mentioned water meter image dataset also includes standard water meter images (i.e., not including challenging scenarios) to ensure a comprehensive evaluation of the model performance under normal and adverse conditions.

[0053] It can be understood that the present invention constructs the above-mentioned water meter image dataset as a structured selection process, integrating manual curation, automatic filtering, and redundancy reduction to obtain a dataset optimized for developing a high-precision recognition model in complex environments.

[0054] S2. Use the water meter image dataset to train the water meter reading recognition model, where the water meter reading recognition model is an improvement based on YOLOv11.

[0055] Specifically, the water meter reading recognition model of the present invention can propose two structural improvements to the existing YOLOv11 n to improve the recognition accuracy in challenging water meter reading scenarios. First, the present invention develops a challenging scenario feature extractor (CSFE) by integrating the AdditiveBlock of CAS-ViT and the convolutional GLU of TransNeXt. CSFE optimizes the original C3k2 module of YOLOv11 n to form the C3k2-CSFE proposed by the present invention (see Figure 4 , upper right corner), significantly enhancing the model's ability to represent features and adapt to environmental changes. Second, the water meter reading recognition model of the present invention adds a deep semantic fusion module PSFM, which adaptively fuses multi-scale features to improve character recognition accuracy and ensure performance consistency under different target sizes and complex backgrounds.

[0056] Vision Transformers (ViTs) capture long-range dependencies through self-attention mechanisms, showing powerful capabilities that enable the model to extract global context information and enhance feature representation. However, the high computational complexity of ViTs increases the processing cost and limits efficiency, especially for tasks that require real-time or resource-efficient computations. CAS-ViT introduces the AdditiveBlock, which reduces the computational overhead while retaining the advantages of ViTs, making the AdditiveBlock more suitable for efficient visual processing. Similarly, TransNeXt introduces Convolutional Gated Linear Units (Convolutional GLU), which combines depth convolution and gating mechanisms to improve feature selectivity and fusion, thereby achieving more effective feature representation. In addition, the present invention also explores feature fusion techniques to enhance the recognition ability under complex imaging conditions. PSFusion is an infrared and visible light image fusion framework designed based on progressive semantic injection and scene fidelity. It introduces a Profound Semantic Fusion Module (PSFM), which refines multi-source feature integration through cross-attention mechanisms and modality-specific attention maps.

[0057] Preferably, the water meter reading recognition model of the present invention includes a Challenging Scene Feature Extractor (CSFE). Among them, as Figure 5 shown, the architecture of CSFE includes three key components: a Local Feature Extraction Component (Integration), a Convolutional Additive Token Mixer (CATM), and a Convolutional GLU (Convolutional GLU) module. The Local Feature Extraction Component enhances spatial feature expression through multi-layer convolution (Conv) and normalization (Norm), capturing the key details required for character recognition. The CATM applies a convolution-based additive attention mechanism, replacing the conventional matrix multiplication and softmax operations in self-attention with spatial and channel interactions. This design significantly reduces the computational complexity while maintaining global context expression. The transformation process in CATM is defined as the following formula (1):

[0058] O = Γ(Φ(Q)+Φ(K))·V (1)

[0059] where Γ represents depth convolution for enhancing context integration, Q is the Query, K is the Key, and V is the Value; the transformation function in formula (1) is defined as the following formulas (2)-(4):

[0060] Φ(X) = Φ Spatial (X)+Φ Channel (X) (2)

[0061] Φ Spatial (X) = σ(W DW (X)) (3)

[0062] Φ Channel (X) = X·σ(W1×1(AvgPool(X))) (4)

[0063] Among them, W DW and W 1×1 represent depth convolution and pointwise convolution respectively, and σ is the activation function. The spatial transformation Φ Spatial (X) enhances local feature extraction, while the channel transformation Φ Channel (X) captures global dependencies through an adaptive attention mechanism.

[0064] Based on the above CATM, the convolutional GLU module dynamically adjusts the importance of features through a gating mechanism to achieve selective feature activation. The convolutional GLU module first projects through a 1×1 convolution and splits into the main feature x and the gating signal v, as shown in the following equation (5):

[0065] (x, v) = Split(W 1×1 (X)) (5)

[0066] The main feature is subjected to depth convolution, as shown in the following equation (6):

[0067] x conv = W DW x (6)

[0068] The gating signal is activated as shown in the following equation (7):

[0069] v = σ(v) (7)

[0070] Among them, represents the allocated channel dimension. The gating signal is calculated as shown in the following equation (8):

[0071] x gated = x conv ·v (8)

[0072] Finally, the convolutional GLU module performs channel fusion on the features through a 1×1 convolution, as shown in the following equation (9):

[0073] x out = W 1×1 xgated (9)

[0074] The convolutional GLU module adds a residual connection to improve the gradient flow and feature reuse, as shown in the following equation (10):

[0075] y = x + x out (10)

[0076] The above design of the challenging scenario feature extractor CSFE of the present invention effectively reduces redundant calculations compared with traditional MLP layers, selectively amplifies information features, and makes it very suitable for robust water meter reading tasks under diverse conditions.

[0077] Preferably, the water meter reading recognition model of the present invention includes a deep semantic fusion module PSFM, which was originally designed for multimodal fusion. On this basis, by integrating high-frequency (F HF ) and low-frequency (F LF ) features to enhance feature expression. F HF captures fine-grained character details crucial for accurate digit recognition, while F LF retains global context stability and improves robustness to environmental changes; by dynamically balancing these components, PSFM enhances multi-scale feature expression and ensures improved water meter reading recognition performance in diverse and challenging scenarios.

[0078] As Figure 6 shown, PSFM includes three main stage operations: feature refinement, cross-frequency attention calculation, and feature fusion. First, for feature refinement, hierarchical features are refined through dense convolutional transformations, which enhance the expression of local structural details while retaining global semantic integrity; given the input feature map F HF , After high-frequency and low-frequency features are processed by dense layers respectively, they are projected into Query (Q), Key (K), and Value (V) spaces, as shown in the following equations (11)-(12):

[0079] Q = Reshape(Conv Q ([F HF , F LF )) (11)

[0080] K, V = Reshape(Conv K (F)), Reshape(Conv V (F)), F ∈ {F HF , F LF} (12)

[0081] Next, for the cross-frequency attention calculation of PSFM, cross-frequency attention is applied to establish the interdependence between F HF and F LF , and a scaled dot-product attention mechanism captures long-range dependencies and achieves effective feature integration, as shown in the following equation (13):

[0082]

[0083] Among them, d k represents the dimension of the key vector. Then, the refined features are calculated as shown in the following equation (14):

[0084] F refined = A·V, A ∈ {A HF , A LF}}, V ∈ {V HF , V LF}} (14)

[0085] Finally, in the feature fusion stage of PSFM, the refined high-frequency and low-frequency components are merged through a convolutional layer to generate the final feature representation, as shown in the following equation (15):

[0086] F out = Conv([F LF，refined , F HF，refined ) (15)

[0087] In addition, although recognition methods play a crucial role in water meter reading, without appropriate evaluation metrics, their effectiveness cannot be comprehensively evaluated. Methods based on character recognition mainly evaluate performance at the single-character level, while existing common object detection metrics such as precision and recall provide insights into single-character detection performance. However, both of these methods lack sequence-level evaluation, which is crucial for ensuring the accuracy of the final water meter reading. Existing sequence-level metrics, including the meter reading rate (MRrate) and the dial recognition rate (DRrate), evaluate the accuracy of the overall sequence but rely on strict exact match requirements, making them overly restrictive when there are minor changes. For example, if the actual reading transitions from 00049 to 00050, a model prediction of either 00049 or 00050 should be considered correct, but MRrate would classify one of them as incorrect. In addition, these metrics do not distinguish errors of different positional importance, and although misclassification of higher-order digits has a greater impact on consumption calculation, they also fail to consider the severity of insertions and deletions in the sequence. Therefore, to comprehensively evaluate the performance of the proposed water meter reading recognition model, the present invention introduces two new evaluation metrics: the Tolerant Table Reading Rate with Small Range Digital Jumps (TTMRrate) and the Weight-Aware Digital Recognition Rate that Values the Accuracy of Higher-Order Digits (WADRrate). These two metrics aim to better reflect the challenges of actual water meter readings by incorporating elements such as digital jump tolerance, sequence length penalty, position-sensitive weighting, and error type differentiation, providing a more comprehensive measure of model performance in water meter reading.

[0088] Specifically, for the Table Read Rate (TTMRrate) that tolerates small - range digital jumps, mechanical water meters often exhibit slight differences during the digital jump process, where the digits partially transition to the next value. These small changes may lead to incorrect evaluations under existing assessment metrics. To account for this inherent imprecision, TTMRrate introduces a tolerance mechanism. If the deviation between the predicted value and the true value does not exceed ±1, the prediction is considered correct. TTMRrate is defined as follows in equations (16) - (17):

[0089]

[0090] where pred m and gt m represent the predicted reading and the true reading of the m - th sample respectively. This TTMRrate metric ensures that slight changes due to mechanical digit jumps do not unduly penalize the model performance, providing a more practical assessment consistent with the actual usage scenario.

[0091] Specifically, for the Weight - Aware Digital Recognition Rate (WADRrate), different from TTMRrate which introduces tolerance for small deviations, WADRrate prioritizes errors based on their position in the sequence, ensuring that higher - order digits contribute more to the final score. This WADRrate metric combines length penalty, position - aware dynamic weighting, and character - level error discrimination to provide a more accurate assessment of recognition performance. In actual water meter reading applications, even if the numerical value of the reading is correct, there may be cases of sequence length mismatch, such as missing or extra leading zeros, which may lead to misunderstandings. To address this issue, WADRrate introduces a length penalty to penalize over - detected or undetected digits, and the penalty calculation is as follows in equation (18):

[0092]

[0093] where len(pred) and len(gt) represent the lengths of the predicted sequence and the true sequence respectively. This rule in equation (18) ensures that models that generate overly long or short predictions are penalized.

[0094] In addition, WADRrate applies a position - sensitive weighting scheme to prioritize higher - order digit errors that have a greater impact on the final water meter reading. The position weight of the i - th digit is calculated as follows in equation (19):

[0095]

[0096] where L represents the length of the longest sequence. Equation (19) ensures that errors in the leading digit are penalized more highly, in line with the actual impact of misreading higher - order digits.

[0097] Furthermore, the character-level recognition score is penalized as follows according to different types of errors: the penalty for substitution errors (incorrect digits) is -w i ·1.0; the penalty for insertion errors (extra digits) is -w i ·1.5; the penalty for deletion errors (missing digits) is w i ·2.0, which is the most severe as it has the greatest impact on the final reading. The total character-level score of a sample is calculated as follows in Equation (20):

[0098]

[0099] where E i represents the error type at position i, and its values are {0: correct, 1: incorrect substitution, 1.5: extra digit, 2: missing digit}. The final WADRrate score is calculated as follows in Equation (21):

[0100]

[0101] The max(0, ·) function ensures that the final score remains non-negative and avoids over-penalization in extreme cases. By combining sequence length penalty, dynamic weighting, and character-specific error differentiation, WADRrate provides a comprehensive evaluation metric that more accurately reflects the impact of recognition errors in actual water meter reading applications.

[0102] S3. Use the trained water meter reading recognition model to perform water meter reading recognition.

[0103] The following conducts an experimental analysis of the model performance of the present invention. Specifically, all experiments can be carried out on a high-performance computing platform equipped with an NVIDIA RTX4090 GPU and an AMD Ryzen 9 7900X CPU. The operating system is Ubuntu 22.04, and PyTorch 2.2.0 and CUDA 12.6 are used. The water meter reading recognition model is trained using the Adam optimizer with a learning rate of 0.01 and a weight decay of 0.0005; the training lasts for 300 epochs with a batch size of 64, balancing computational efficiency and stability.

[0104] For evaluation, the present invention adopts single-character and sequence-based metrics. The single-character evaluation includes precision (P), recall (R), and mean average precision (mAP), and mAP@50 and mAP@50:95 are used to evaluate the detection performance at different IoU thresholds. The sequence-based evaluation includes the following established metrics (1)-(3):

[0105] (1) Meter Reading Rate (MRrate): This metric measures the proportion of sequences where the predicted reading exactly matches the ground truth, and its calculation formula is as follows in Equation (22):

[0106]

[0107] Among them, match(x, y) returns 1 when x=y, otherwise it returns 0.

[0108] (2) Dialing Recognition Rate (DRrate): This metric uses the Levenshtein distance to evaluate character-level accuracy. The Levenshtein distance measures the minimum number of edits required to convert a predicted sequence into a true sequence. Its formula is as follows (23):

[0109]

[0110] Among them, lev(pred m ,gt m ) is the Levenshtein distance between the predicted and true series.

[0111] (3) Mean absolute error (MAerror): This indicator calculates the absolute difference between the predicted reading and the true reading, providing a measure of numerical deviation. Its formula is as follows (24):

[0112]

[0113] Among them, p m and g m The predicted readings and the actual readings respectively.

[0114] In addition, the present invention also adopts the TTMRrate and WADRrate evaluation indicators as mentioned above to ensure the practicality of model evaluation. In the ablation study, only single-character evaluation indicators are used to evaluate the impact of individual components on object detection performance. In contrast, in the comparative experiments, both single-character and sequence-based metrics are used to comprehensively evaluate the effectiveness of the model in real-world scenarios.

[0115] Experimental results and analysis:

[0116] First, an ablation study was conducted to verify the impact of the C3k2-CSFE and PSFM modules of the present invention. As shown in Table 1 below, each module significantly contributes to the performance improvement of the baseline YOLOv11 n. After combining these two modules, the precision (P) and recall (R) reached 0.957, and the mAP@50 and mAP@50:95 were improved to 0.983 and 0.861 respectively; the above findings confirm that the model enhancement measures proposed in the present invention effectively enhance feature expression and positioning accuracy.

[0117] Table 1 Ablation study results

[0118]

[0119] To further evaluate the effectiveness of the water meter reading recognition model of the present invention in actual scenarios, the present invention conducted comparative experiments using single-character and sequence-based metrics. Table 2 below shows the results of single-character recognition. The model of the present invention outperforms all competing YOLO models in all metrics; compared with the best-performing baseline YOLOv8n, the method of the present invention improved by 2.9% in precision (P), 1.7% in recall (R), and 4.6% in mAP@50:95. Notably, the model of the present invention achieved an mAP@50:95 of 0.861, exceeding all baselines and demonstrating its outstanding accuracy in high-precision scenarios.

[0120] Table 2 Comparison Table of Single-Character Evaluation

[0121]

[0122] Table 3 below shows the sequence-based evaluation results, and the model of the present invention consistently achieved the highest performance in all metrics. The method of the present invention reduced the MAerror to 11,305.17, a reduction of 83.5% compared to the best-performing baseline, significantly enhancing the accuracy of numerical prediction. In addition, the WADRrate of the present invention reached 92.68%, exceeding all competing models. This indicates that the impact of prediction errors on the final water meter reading is minimal, as the model of the present invention effectively alleviates problems such as missing digits, extra digits, and misclassification of high-order digits. In addition, the model of the present invention obtained the highest TTMRrate of 96.19%, confirming its outstanding ability to identify semi-characters caused by mechanical digit jumps.

[0123] Table 3 Comparison Table of Sequence-Based Evaluation

[0124]

[0125]

[0126] To further verify the effectiveness of the model of the present invention, a qualitative analysis of the detection results under various challenging actual conditions was conducted. Figure 7 It demonstrates the outstanding adaptability of the model in diverse adversities such as extreme light changes, water droplet occlusion, dial tilt, and other environmental interferences. Despite these severe challenges, the model of the present invention ( Figure 7 Ours) still showed outstanding accuracy in detecting and classifying digits. This performance is in sharp contrast to the baseline YOLOv11 n, which showed obvious limitations of frequent missed detections - caused by Figure 7The red arrows and dashed boxes highlight the undetected numbers. Comparative analysis not only reveals the deficiencies of the baseline model but also emphasizes the advantages of the model proposed in this invention in enhancing detection capabilities. The model of this invention can maintain consistent and accurate performance even under adverse conditions, confirming its robustness and highlighting its potential to significantly improve recognition accuracy in complex water meter reading environments.

[0127] In addition to improving robustness under adverse conditions, the water meter reading recognition model of this invention also significantly improves the accuracy of character recognition, such as Figure 8 shown. A key improvement of this invention is its ability to accurately distinguish visually similar numbers, which is a common source of misclassification in conventional models; for example, due to structural similarity and lighting changes, YOLOv11 n often misidentifies '9' as '3', while the model of this invention ( Figure 8 Ours) correctly identifies these numbers, demonstrating better feature extraction and classification accuracy. In addition, the model of this invention also demonstrates a strong ability to recognize semi-characters caused by mechanical digit jumps. Different from the baseline YOLOv11 n, which has difficulty recognizing these transitional states, the model of this invention effectively detects and interprets these ambiguous characters, achieving the highest TTMRrate among all baselines.

[0128] Further analysis shows that the model of this invention can well adapt to various water meter designs, including differences in font styles, digit spacing, and protective covers, ensuring high recognition accuracy in diverse environments; this versatility is crucial for practical deployment because water meters may vary significantly in appearance and conditions. The improvements in adaptability, robustness, and precision together enhance the effectiveness of the method of this invention, highlighting its potential to significantly improve water meter reading accuracy in real-world applications.

[0129] It is understandable that accurate water meter readings are crucial for fair billing and efficient water resource management. However, achieving high recognition accuracy remains challenging due to issues such as degraded image quality caused by harsh environmental conditions and half-characters generated by mechanical digit jumps. To address these challenges, the present invention proposes a comprehensive solution. The method of the present invention systematically integrates the planning of the dataset, the enhancement of the model, and the improvement of the evaluation, forming a unified strategy to improve the recognition performance. By solving the environmental factors and mechanical digit jump problems, the accuracy and robustness of water meter reading recognition under challenging conditions are improved. First, the present invention plans a water meter image dataset including diverse challenging scenarios such as extreme lighting changes, accumulated stains, surface wear, and water droplets, systematically incorporating various challenging conditions, ensuring comprehensive coverage of challenging environmental conditions, ensuring robust learning of challenging samples and enhancing the adaptability of the model, laying a solid foundation for model training and generalization. Second, using the above water meter image dataset, the present invention introduces an enhanced YOLOv11 framework. The water meter reading recognition model of the present invention integrates CSFE and PSFM. CSFE integrates the AdditiveBlock of CAS-ViT and the convolutional GLU of TransNeXt, improving feature extraction and representation; PSFM optimizes multi-scale feature fusion and enhances the robustness to mechanical digit jumps and adverse environmental conditions; the water meter reading recognition model of the present invention optimizes the YOLOv11 n framework and enhances the recognition robustness of YOLOv11 n, significantly improving the recognition robustness in challenging scenarios. Finally, the present invention proposes two new metrics, TTMRrate and WADRrate, for comprehensive and practical performance evaluation. TTMRrate can tolerate minor digit jump errors to reduce unnecessary penalties; while WADRrate prioritizes the accuracy of higher-order digits and penalizes errors based on position-aware weighting and error type differentiation; TTMRrate and WADRrate provide a more practical and accurate evaluation of the recognition performance of the water meter reading recognition model. Experimental results show that the method of the present invention is effective, achieving an mAP@50 of 98.3% and a TTMRrate of 96.19%, exceeding the performance of existing YOLO-based models in single-character and sequence-level recognition tasks.

[0130] It should be noted that the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, such that a process, method, article or apparatus comprising a series of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article or apparatus. Without further limitation, an element defined by the phrase "comprising an..." does not exclude the presence of additional identical elements in the process, method, article or apparatus comprising said element.

[0131] Although the embodiments of the present invention have been shown and described, those of ordinary skill in the art can understand that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A robust water meter reading recognition method in challenging scenarios based on YOLOv11, characterized in that, Including the following steps: S1. Construct a water meter image dataset, where the water meter image dataset includes water meter images of various different challenging scenarios; S2. Use the water meter image dataset to train a water meter reading recognition model, where the water meter reading recognition model is an improvement based on YOLOv11; S3. Use the trained water meter reading recognition model to recognize water meter readings.

2. The robust water meter reading recognition method in challenging scenarios based on YOLOv11 according to claim 1, characterized in that: The challenging scenarios include overexposure, underexposure, glare, stains, surface wear, and water droplets.

3. The robust water meter reading recognition method in challenging scenarios based on YOLOv11 according to claim 1, characterized in that: The water meter image dataset also includes standard water meter images.

4. The robust water meter reading recognition method in challenging scenarios based on YOLOv11 according to claim 1, wherein The step S1 includes the following sub-steps: S11. Collect multiple original water meter images; S12. Select multiple water meter images of different challenging scenarios from the multiple original water meter images as training samples; S13. Use the training samples to train a lightweight classification network; S14. Use the lightweight classification network to perform rough screening on the multiple original water meter images to obtain a roughly screened water meter image dataset; S15. Eliminate redundant images in the roughly screened water meter image dataset; S16. Manually re-screen to construct the water meter image dataset.

5. The robust water meter reading recognition method in challenging scenarios based on YOLOv11 according to claim 1, characterized in that, Before the step S2, it also includes: performing word wheel digit marking on the water meter images in the training set and validation set of the water meter image dataset, where the word wheel digits include 20 different categories: 10 integer digit categories and 10 semi-character categories representing intermediate transition states.

6. The robust water meter reading recognition method in challenging scenarios based on YOLOv11 according to claim 1, characterized in that: The water meter reading recognition model includes a challenging scenario feature extractor CSFE, where the CSFE includes a local feature extraction component, a convolutional addition token mixer CATM, and a convolutional GLU module.

7. The robust water meter reading recognition method in challenging scenarios based on YOLOv11 according to claim 1, characterized in that: The water meter reading recognition model includes a deep semantic fusion module PSFM, where the PSFM includes feature refinement, cross-frequency attention calculation, and feature fusion.

8. The robust water meter reading recognition method in challenging scenarios based on YOLOv11 according to claim 1, characterized in that, It also includes: Introducing a meter reading rate that tolerates small-range digital jumps to evaluate the performance of the water meter reading recognition model.

9. The robust water meter reading recognition method in challenging scenarios based on YOLOv11 according to claim 1, characterized in that It also includes: Introducing a weight-aware digital recognition rate to evaluate the performance of the water meter reading recognition model.

Images