Inter national J our nal of Recongurable and Embedded Systems (IJRES) V ol. 15, No. 3, No v ember 2026, pp. 629 ∼ 644 ISSN: 2089-4864, DOI: 10.11591/ijres.v15.i3.pp629-644 ❒ 629 Image pr ocessing and computer vision on lo w-cost micr ocontr ollers: a case study on water meter r eading Sohrab J alali 1 , Filiz Ers ¨ oz 2 , Babak T aeb 3 1 Department of Softw are Engineering, Graduate School of Natural and Applied Sciences, OST ˙ IM T echnical Uni v ersity , Ankara, T urk e y 2 Department of Data Science and Analytics, F aculty of Economics and Administrati v e Sciences, OST ˙ IM T echnical Uni v ersity , Ankara, T urk e y 3 Independent Researcher , Ankara, T urk e y Article Inf o Article history: Recei v ed Jun 16, 2026 Re vised Aug 14, 2026 Accepted Sep 11, 2026 K eyw ords: Adapti v e sampling Anomaly detection Embedded computer vision ESP32-CAM INT8 quantization T in yML W ater meter reading ABSTRA CT Despite the w orldwide mo v e to w ards smart meters, mechanical w ater meters are still used e xtensi v ely in residential and municipal netw orks. Upgrading them with computer vision a v oids cost ly infrastructure replacement, yet most approaches send data to cloud serv ers for processing. This paper introduces an end-to-end on-de vice solution for recognizing w ater meter digits on the ESP32- CAM board, which costs belo w $15 and features only 520 KB SRAM. A MobileNetV2-based digit classier is quantized to INT8. Three rmw are im- pro v ements e xtend the inference loop: i) a temporal check conrming a reading only when three consecuti v e detections match, ii) on-de vice anomaly detection without serv er support, and iii) adapti v e sampling between 3 s during w at er us- age and 60 s when idle. A Flask-based internet of things (IoT) dashboard named MeterW atch pro vides li v e monitoring. The model achie v es 2 ms latenc y and 235.6 KB peak memory . On a 553-image test subset from a custom 6,155-image dataset with 13,470 digit annotations it reaches 99.60% cell-le v el accurac y in- cluding background, 96.89% digi t-only accurac y and 87.70% e xact v e-digit reading accurac y . T o the best of our kno wledge, this is the rst sys tem combin- ing on-de vice INT8 deplo yment, w a ter meter digit recognition and intelligent rmw are beha viour on the ESP32-CAM. Ev aluations used a single meter brand under x ed ash illumination. This is an open access article under the CC BY -SA license . Corresponding A uthor: Sohrab Jalali Department of Softw are Engineering, Graduate School of Natural and Applied Sciences OST ˙ IM T echnical Uni v ersity Ankara, T urk e y Email: sohrabjallali@gmail.com 1. INTR ODUCTION Edge articial intelligence (AI) in conjunction with cheap embedded de vices has made a ne w form of intelligence possible to deplo y . One of the most adv anced de vices in this cate gory is the ESP32-CAM which comprises a dual-core system on chip (SoC) de vice along with an O V2640 camera and 802.11 b/g/n W i-Fi capability . W ith its cheap price of less than $15 per unit, this platform represents a good replacement for ener gy intensi v e s ingle-board computers lik e Raspberry Pi and NVIDIA Jetson for massi v e sensing purposes. It can e v en be used inside already-e xisting enclosures at city scale. One area that is poised to see great g ains is w ater utilities. The traditional mechanical meter is by f ar the most common de vice for measuring w ater consumption in both pri v ate homes and city w ater systems J ournal homepage: http://ijr es.iaescor e .com Evaluation Warning : The document was created with Spire.PDF for Python.
630 ❒ ISSN: 2089-4864 around the w orld. Ho we v er , the y still require manual meter reading in almost all cases. This procedure is both inef cient and costly not to mention that it is subject to e rror from human transcription [1]. Upgrading all of these meters to contemporary digital meters is a nancially prohibit i v e proposition, costing man y millions of dollars. A less costly solution in v olv es retrotting these meters with computer vision capabilities that capture and interpret the digit display in real time [2], [3]. Implementing deep learning on a microcontroller presents se v eral challenges. The re gular object detection algorithms consume hundreds of me g abytes of memory , alongside the need for GPU or NEON optimization, neither of which is a v ailable on the ESP32 board. Apart from the memory requirements, the tw o core processors on the ESP32 w ork at a maximum speed of 240 MHz, which is enough for an embedded de vice application b ut w ay slo wer than the computing po wer required for a typical inference algorithm. Achie ving practical accurac y on real-w orld meter images which e xhi b i t v ariable lighting, angular distortion, glare, dirt, and surf ace wear needs both careful model design and rigorous compression [4], [5]. The e v aluation reported her e isolates the modeling and compression problem under a controlled x ed-illumination deplo yment. Generalisation to uncontrolled lighting remains future w ork. The proposed paper pro vides an inte grated approach to w ard solving the abo v e problem . Some of the main points include the follo wing: - An INT8-quantized MobileNetV2-based digit classier running enti rely on the ESP32-CAM o v er a x ed 12 × 12 grid, achie ving 2 ms model inference latenc y , 235.6 KB peak RAM, a 99.60% o v erall cell-le v el accurac y including the background class, and a 96.89% digit-only classication accurac y on a 553-image box-annotated test subset. - A temporal consistenc y technique where the rolling b uf fer consists of three readings and an y reading is considered v alidated only if three consecuti v e readings match, ltering transient errors and sensor noise. - On-de vice anomaly detection using rmw are-le v el arithmetic on conrmed readings to ag abnormal consumption e v ents such as pipe leaks, without an y serv er in v olv ement. - Adapti v e sampling: a time-interv al-based system that changes the interv al length from 60 seconds when the system is idle to 3 seconds when there is an y acti vity . It helps in sa ving c ycles of inference. - MeterW atch: an internet of things (IoT) dashboard de v eloped using Flask technology which supports li v e JPEG images with bounding box es, trend graphs using Chart.js, anomalies, multi-de vice capability , and session statistics. - A custom dataset of 6,155 w ater meter images, including 2,694 images with digit-le v el bounding box annotations totalling 13,470 box es, obtained under real-w orld scenarios. 2. RELA TED W ORK 2.1. W ater meter and utility meter r eading Dif ferent approaches ha v e been tak en to research automatic w ater meter reading. Initially , approaches were based on traditional image processing techniques such as thresholding, morphological processing, and template matching. Ho we v er , these approaches suf fer from poor rob ustness due to f actors such as v arying lighting conditions and de graded surf aces [1]. Deep l earning models pro v ed more ef fecti v e. In particular , Nguyen et al. [2] utilized te xt recognition techniques along with a deep learning approach. Their model demon- strates competiti v e performance when applied to a lar ge dataset of meters. The only dra wback is that the system uses a GPU serv er for inference. Zhang et al. [3] implemented an attention model at a character le v el in order to o v ercome partially occluded or blurred digits. The y also apply inference on a GPU serv er , which is used on cloud computing platforms. Hong et al. [4] tested their meter reader model in deplo yment settings character - ized by hea vy fouling, lo w lighting, and se v ere vie wing angles and found that data augmentation signicantly enhances accurac y . Their system is e x ecuted on w orkstations. Li et al. [5] focused on mechanical w ater meters and reported 97.9% mAP@0.5 with an impro v ed Y OLOv4 detector in natural scenes without the need for em- bedded deplo yment. Martinelli et al. [6] incorporated meter reading into a lar ger smart grid monitoring system and used cloud computing infrastructure for inference. Qiao et al. [7] de v eloped a compact v ariant of Y OLO named Meter -Y OLOv8n that supports w ord-wheel w ater meters b ut still requires serv er -side deplo yment. A major limi tation in the surv e y arises in this conte xt: se v en prior studies specically focus on the task of digit recognition in w ater meters, of which only one (Ne v e [8]) tar gets the ESP32-CAM. Ne v ertheless, the w ork of Ne v e [8] lacks INT8 quantization, an y memory proling, or anomaly and adapti v e capabilities. This limitation stems from a technical perspecti v e. First, due to the limited RAM of the ESP32 (520 KB), the stan- Int J Recongurable & Embedded Syst, V ol. 15, No. 3, No v ember 2026: 629–644 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Recongurable & Embedded Syst ISSN: 2089-4864 ❒ 631 dard format of a machine learning model cannot be used, second, due to the requirement of proper calibration in case of TFLite Micro, and third, due to the need for an annotated dataset suitable for the tar get meter and deplo yment conditions. Be yond these g aps, Ne v e’ s system relies on manually dened re gions of interest for each digit position, congured once through a calibration interf ace and x ed thereafter , rather than a learned localisation or classication mechanism. The captured images are also de scribed as being sent o v er W i-Fi for processing, with the actual location of inference (on-de vice or serv er -side) left unspecied. No accurac y , mem- ory footprint, or latenc y gures are reported, which limits direct quantitati v e comparison (see subsection 4.7). Our w ork addresses precisely this g ap. T able 1 summarises the comparison. T able 1. W ater meter reading studies: platform summary (7 prior w orks and this w ork, 6 of the 7 prior w orks rely on serv er/cloud inference) Reference Platform On-de vice INT8 Ne v e [8] ESP32-CAM Unspecied No Nguyen et al. [2] DL serv er No N/A Zhang et al. [3] CNN serv er No N/A Hong et al. [4] ML serv er No N/A Li et al. [5] CNN serv er No N/A Martinelli et al. [6] IoT + cloud No N/A Qiao et al. [7] Y OLO serv er No N/A This w ork ESP32-CAM Y es Y es 2.2. On-de vice infer ence on ESP32 There are some papers related to on-chip inference in ESP32. Ho we v er , v ery fe w w orks succe ssfully deplo y an end-to-end embedded solution. The research paper most rele v ant to our case is that of Ben Aicha [9], who conducts INT8-quantized CNN inference on ESP32-CAM for generic image classi cation. The paper sho ws ho w ef fecti v e the INT8 post-training quantization is in impro ving the accurac y of the model after the training process. It pro v es the possibil ity of performing real-time embedded inference under the memory limitation of the ESP32. Nonetheless, the paper does not consider the cases of w ater -metering reading and anomaly detection nor adapti v e system beha vior . Sine glazo v and Khotsyano vsk y [10] sho wed that the implementation of the entire con v olutional neural netw ork architecture w as possible right on the ESP32 without an y pre-processing steps performed e xternally . The w ork of Dokic [11] took a simpler approach through using logistic re gression on lo w-resolution images captured by the ESP32 camer a. On the other hand, P aul et al. [12] applied re gression algorithms in embedded systems for the task of reading analog pressure g auges in a laboratory en vironment. Ho we v er , current research using ESP32-CAM mostly in v olv es taking the pictures locally , after which inference is done of f-de vice. According to Narw aria et al. [13] and Sulistiyo w ati et al. [14], local pic- ture captur e is follo wed by sending the captured images to a PC for processing with Y OLO or OpenCV . Hadz et al. [15] managed to int e grat e Y OLO with ESP32-CAM for stereoscopic depth estimation, where inference is carried out of f-de vice too. In addition, No wroz et al. [16] and Im ron et al. [17] conducted infer - ence on a cloud-based back-end system. Setia w an et al. [18] combined Y OLOv8 with ESP32-CAM capture b ut processed detections serv er -side. Chang et al. [19] carried out object detection using ESP32 edge computing and Ne vliudo v et al. [20] conducte d object recognition with a microcontroller -based robot. The aforemen- tioned of f-de vice processing represe n t s the most widely used methodology in ESP32-CAM research, whereas the present research represents only one of the fe w e xceptions from that rule. 2.3. T inyML and model compr ession The basis of theory for INT8 quantization is laid by Nagel et al. [21] as the y analyze the roots of quantization error and de v elop calibration methods to counteract s u c h ne g ati v e ef fects. Theory into practice w as achie v ed through research done by No v ac et al. [22] whereby the y implemented quantized deep neural netw orks on Corte x-M microcontrollers and analyzed the trade-of f curv e of accurac y v ersus memory footprint at dif ferent bit-widths. Finally , Shabir et al. [23] conduct an analysis comparing pruning ag ainst quantization methods on T in yML applications concluding that post-training INT8 quantization i s optimal for accurac y and ef cienc y trade-of f. Ima g e pr ocessing and computer vision on low-cost micr ocontr oller s: a case study on ... (Sohr ab J alali) Evaluation Warning : The document was created with Spire.PDF for Python.
632 ❒ ISSN: 2089-4864 T in yissimoY OLO w as introduced by Moosmann et al. [24], which is a fully quantized Y OLO v ariant designed specical ly for ultra-lo w-memory MCUs, reporting sub-1 MB memory operation on ARM Corte x-M class hardw are. Further progress in this direction w as made by Bo yle et al. [25] with DSOR T -MCU, focusing on de v eloping methods that w ould enable small object detection on microcontroller units. In addition, Zhang [26] e xamined quantization-a w are training (QA T) as a substitute approach compared to post-training quantization in embedded articial intelligence systems. MC UNetV2 w as de v eloped by Lin et al. [27], where inference w as done using patches in order to reduce peak acti v ation memory requirements do wn to 256 KB. 2.4. Embedded inter net of things and anomaly detection Ba w ankar et al. [28] de v eloped a retrot w ater -monitoring system that recognizes meter digits from camera images on a Raspberry Pi Zero W edge de vice and transmits the readings to the cloud for further analysis. Our anomaly detection algorithm is quit e unique in nature since it detects abnormalities through vision-based meter readings and completes all the processing logic via ESP32 rmw are. P a w ar et al. [29] con- ducted a surv e y on smart ener gy meters using ESP32 to justify its applicability in utility IoT . In addition, Samanta et al. [30] used T in yML in combination with ESP32-CAM for disaster location identication in the real w orld pro ving the ef fecti v eness of on-de vice inference under real deplo yment constraints. Lastly , Donsk o y et al. [31] adopted T in yML on the ESP32 for agricultural classication to sho w that embedded mod- els can maintain reliable accurac y across di v erse real-w orld en vironments. 3. METHOD 3.1. System o v er view Figure 1 sho ws the pipeline o v ervie w . In the rst step, the ESP32-CAM captures a JPEG image and then performs its pre-processing and INT8 inference completely onboard before applying the three rmw are contrib utions and sending the output data to MeterW atch via W i-Fi. The pre-processing procedure, which in v olv es trans forming from 320 × 240 JPEG to 96 × 96 graysca le, is done on-de vice by decoding the JPEG format and performing bilinear do wnsampling, adding approximately 8–12 ms to the pipeline latenc y . Model inference itself accounts for only 2 ms. Only the nal POST request lea v es the de vice. Figure 1. End-to-end system pipeline with all inference and rmw are logic e x ecuted on the ESP32-CAM 3.2. Hard war e platf orm The ESP32-CAM board runs on Espressif ’ s ESP32 chip, which has tw o Xtensa LX6 cores that can hit speeds up to 240 MHz. It packs 520 KB of SRAM, 4 MB of PSRAM and 4 MB of Flash memory , along with a b uilt-in O V2640 camera. W i-Fi 802.11 b/g/n is supported nati v ely without additional hardw are. Priced at under $15 per unit, it is an af fordable option for lar ge-scale infrastructure deplo yment. Int J Recongurable & Embedded Syst, V ol. 15, No. 3, No v ember 2026: 629–644 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Recongurable & Embedded Syst ISSN: 2089-4864 ❒ 633 3.3. Dataset collection and annotation The images were captured using a self-designed and manuf actured 3D-printed data acquisition de vi ce (Figure 2). The de vice consists of a rigid structure that includes a base with rail on which the re are four v ertically mounted guide rods to allo w mo v ement of the camera mount in all three dimensions. While v ertical adjustment is done by mo ving the camera mount up and do wn the rods, independent horizontal adjustments can be made by v arying the positions of thumb-scre w-mounted cameras. This e xibility enabled systematic capture of images across a controlled range of distances and angles. Figure 2. Custom 3D-printed data acquisition rig with a camera mount repositionable in all three ax es A customized dataset w as created using the image capturing of the Baylan-brand mechanical w ater meter across a range of camera-to-meter distances and observ ation angles up to ∼ 30 ◦ of f-axis, with some im- ages e xhibiting surf ace de gradation including glare, dirt, and scratches. Illumination w as held constant across all captures via the rig’ s inte grated ash LED, matchi ng the illumination used in the production enclosure. Only distance and vie wing angle v aried systematically . All images were captured at 320 × 240 pix els in JPEG format. Deplo yment enclosure: as for the production deplo yment, the ESP32-CAM is housed in a sealed 3D-printed enclosure that sits directly o v er the f ace of the meter . This ensures that an y form of ambient light interference is elim inated while pro viding a consistent illumination en vironment through the inte grated ash LED of the O V2640 sensor . Controlled illumination is one of the k e y f actors responsible for the high reported accurac y (99.60%), as the model operates under conditions closely matching those of data collection: no uncontrolled light sources, since the ash-illuminated capture re gime is identical between data collection and production. The cam era-to-meter distance and vie wing angle are x ed at 8 cm and 0 de grees of f-axis specically in the production enclosure, whereas the training and test data span the wider 5–12 cm / 0–30 de gree range listed i n T able 2, since the train-test split w as performed at image le v el rather than by condition (see subsection 4.1). T able 2. Data collection rig parameters P arameter V alue/range Camera-to-meter distance 5–12 cm (production: 8 cm) Horizontal vie wing angle 0°–30° of f-axis V ertical tilt angle 0°–15° of f-axis Illumination source Inte grated O V2640 ash LED Capture resolution 320 × 240 JPEG Meter brand Baylan mechanical (single) Ima g e pr ocessing and computer vision on low-cost micr ocontr oller s: a case study on ... (Sohr ab J alali) Evaluation Warning : The document was created with Spire.PDF for Python.
634 ❒ ISSN: 2089-4864 T otal number of images in this dataset is 6,155, where 4,961 of them represent the training set while 1,194 images form the testing set (80%/20% ratio). Of these, 2,694 images carry digit-le v el bounding box annotations (2,141 train, 553 test), each with all v e digits indi vidually mark ed, yielding a total of 13,470 bounding box annotations. Of these annotations, 9,622 were hand-dra wn and 3,848 were generated via an au- tomatic suggestion mechanism based on a hand-dra wn reference box in the same image, subsequently re vie wed and corrected by hand. The remaining 3,461 images (2,820 train, 641 test) contain no digit-le v el box es and no ground-truth label of an y kind, including no full meter reading v alue. The label eld for these les carries only a le identier . Consequently these images were not used in model training: the ef fecti v e training sample size is limited to the 2,141 box-annotated training images, for which cell-le v el ground truth could be generated. The 641 box-unlabelled test images were run through inference after training solely to observ e whether the model produced a structurally complete v e-digit output. In the absence of ground truth this cannot be scored for correctness. 551 of the 641 images (86.0%) yielded a complete v e-digit output group, which is a structural- completeness indicator only and not a reading-accurac y metric. Digit widths range from 25 to 41 pix els and heights from 37 to 61 pix els. The distrib ution of the number of instances per class sho ws an imbalance, where the most common class is “3” with 1,965 instances (14.6%), and the least common are “1” and “6” with 1,158 instances each (8.6%). Pre-processing steps applied before training included centre-cropping to the digit re gion, grayscale con v ersion, and normalisation to a 96 × 96 pix el input. Augmentation included random rotation ( ± 15 ◦ ), zoom (0.85–1.15 × ), and Gaussian blur . 3.4. Model ar chitectur e selection The choice of an architecture w as limited to the f act t hat the ESP32 only has 520 KB of SRAM mem- ory . Most OCR librarie s (such as T esseract or EasyOCR) and object d e tection algorithms (including Y OLO, F aster R-CNN, or v ariations of SSDs) consume hundreds of me g abytes of memory , and require x86 or ARM processors with dedicated NEON or CUD A acceleration, hardw are entirely absent on the ESP32. In the conte xt of suitable architectures for embedded de vi ces, MobileNetV2 w as selected because depthwise separable con v o- lutions allo w for fe wer multiply-accumulate operations while still pro viding suf cient representational capac- ity to recognize digits among ten classes. Compared to MobileNetV1, MobileNetV2 uses an in v erted residual structure, which yields better gradient o w and lo wer acti v ation memory at equi v alent parameter counts. The nal INT8-quantized model ts within 235.6 KB peak RAM, less than half the a v ailable b udget. 3.5. Model ar chitectur e and training The digit recognition model is made up of a backbone inspired by MobileNetV2’ s in v er ted-residual block structure, signicantly narro wed relati v e to the standard architecture, combined with a dense classica- tion head, accepting 96 × 96 grayscale input. Depthwise separable con v olutions k eep the multiply-accumulate count and parameter total within the ESP32’ s memory b udget. The backbone reduces the input spatially to a x ed 12 × 12 grid of cells, each co v ering an 8 × 8-pix el re gion of the input. Each cell is independently classied via softmax into ele v en cate gories: background or one of ten digit classes. No bounding box coordinates are re gressed. At inference, adjacent cells sharing the same predicted digit class abo v e the condence threshold are geometrically mer ged into a single bounding box for display purposes, without an y learned localisation component. This mer gi ng follo ws a simple axis-aligned-box (AABB) adjacenc y test rather than an intersection o v er union (IoU) threshold: the grid is scanned, each positi v e cell starts or e xtends a group, and cells are joined to a group when their boundaries touch or o v erlap, with a nal mer ge pass across all groups. Cells of dif ferent classes ne v er mer ge into the same group because each cell carries a single softmax-ar gmax label. T raining labels were generated from the digit-le v el bounding box es described in subsection 3.3 as follo ws: the centre (cx, c y) of each b ox w as computed and mapped to a single grid cell via oor di vision by the 8-pix el cell stride, cell x = oor(cx/8) and cell y = oor(c y/8). Only that single centre cell w as assigned the corresponding digit label. Other cells geometrically o v erlapped by the same box were not separately labelled. All remaining cells def aulted to the background class with no additional rule. No IoU-bas ed o v erlap threshold w as used. The cell-label ass ignment depends solely on the centre point. Because adjacent digits in a v e-digit reading are spaced well be yond one cell width apart, no tw o digits’ centres were observ ed to f all within the same cell in this dataset. Int J Recongurable & Embedded Syst, V ol. 15, No. 3, No v ember 2026: 629–644 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Recongurable & Embedded Syst ISSN: 2089-4864 ❒ 635 T raining w as carried out using cloud GPU infrastructure without ImageNet pretraining. All la y e rs were trai ned from scratch, starting from a learning rate of 1e-4, wi th ReduceLR OnPlateau (patience=5, f ac- tor=0.5) halving the learning rate whene v er v alidation loss f ailed to impro v e for v e epochs. T raining w as bounded by EarlyStopping (patience=10, restore best weights=T rue) with a maximum of 1,000 epochs and a batch size of 32. The loss function is a per -cell weighted cate gorical cross-entrop y , with the weight for each digit class computed as a x ed fore ground boost f actor (15.0) multiplied by the ratio of the median class fre- quenc y to that class’ s frequenc y . The background class weight is x ed at 1.0. A x ed random seed (1337) w as used for the train/v alidation split and data shuf ing. No seed w as x ed for in-training-loop randomness such as weight initialisation or batch ordering, so the training procedure itself is not fully deterministic. The de v elopment en vironment used T ensorFlo w 2.21.0, though the e xact v ersion used for the original training run could not be conrmed from the code base. After oat-precision training, a dedicated calibration subset of 200 images, collected separately from the training set and not dra wn from the training split, w as used to deri v e INT8 acti v ation ranges via representati v e dataset calibration. 3.6. INT8 quantization Post-training INT8 quantization con v erts all we ights and act i v ations from FP32 to 8-bit inte ger rep- resentation, follo wi ng Nagel et al. [21] as v alidated for MCU deplo yment by No v ac et al. [22]. The Float32 model w as ne v er e x ecuted on the ESP32. At 887.2 KB it e xceeds the board’ s 520 KB SRAM b udget. Its 98.54% accurac y gure w as obtained from a generic e v aluation-function output in the training en vironment (PC/GPU, T ensorFl o w), not from the detailed 553-image box-annotated confusion-matrix analys is used for the INT8 model in subsection 4.2. Because the tw o gures deri v e from dif ferent e v aluation methodologies, no class-le v el F1-score is a v ailable for Float32 and the tw o accurac y v alues are not directly comparable. The com- piled INT8 TFLite binary occupies approximately 142 KB (141,952 bytes) on disk as a atb uf fer . The EON compiler con v erts this into ra w C++ arrays and direct function calls, stripping the atb uf fer schema, reection, and operator -code-table o v erhead, which is wh y the resulting 94.8 KB on-de vice binary is smaller than the source atb uf fer le. T able 3 summarises the comparison. T able 3. Float32 vs. INT8 model: k e y metrics Metric Float32 INT8 Change Peak RAM (KB) 887.2 235.6 3 . 8 × ↓ Flash (KB) — 94.8 — Inference (ms) — 2.04 (min 1.98, max 2.17, SD 0.05, n=100) — Accurac y (%) 98.54 99.60 +1.06 pp ∗ F1-score — 0.9355 — ∗ Ra w arithmetic dif ference between v alues obtained under dif ferent e v aluation methodologies (see te xt). Not tested for statistical signicance and should not be read as e vidence that INT8 impro v es accurac y . 3.7. Firmwar e contrib utions Three mechanisms were added to the ESP32 rmw are on top of the basic inference loop. 3.7.1. T emporal consistency check The rmw are maintains a rolling b uf fer of the three most recent successful readings. A reading is la- belled conrmed only when all three entries are identical. This eliminates single-frame errors, sensor noise, and motion artef acts. Each transmission to the serv er carries a confirmed ag so the dashboard can distinguish v eried readings from tentati v e ones. 3.7.2. On-de vice anomaly detection The rmw are maintains a reference v alue that is set at the last reset point, either the pre vious anomal y e v ent or the conrmed reading at session start, and does not update on e v ery reading. After each conrmed reading, the rmw are comput es the cumulati v e dif ference between the ne w v alue and this x ed reference. If this dif ference e xceeds a congurable threshold (def ault: ten units), the reading is classied as an anomaly , the anomaly a g is set in the transmission payload, and the reference v alue is reset to the current reading. Because the reference only resets on an anomaly e v ent rather than e v ery c ycle, the same rule catches both sudden usage spik es and slo w leaks that accumulate gradually across man y reading c ycles. This enables detection of pipe leaks or une xpected high-usage e v ents entirely on the microcontroller , without an y serv er -side in v olv ement. Ima g e pr ocessing and computer vision on low-cost micr ocontr oller s: a case study on ... (Sohr ab J alali) Evaluation Warning : The document was created with Spire.PDF for Python.
636 ❒ ISSN: 2089-4864 The def ault threshold of ten units w as selected empirically based on the maximum obser v ed cumulati v e con- sumption increme nt during normal usage on the test meter . It is e xposed as a compile-time constant to allo w operators to calibrate it for meters with dif ferent o w proles or usage patterns. The approach is fundamentally dif ferent from Ba w ankar et al. [28], who transmit vision-deri v ed readings to the cloud for further analysis. 3.7.3. Adapti v e sampling The reading interv al is adjusted dynamically . When the units digit of the meter v alue stays unchanged across consecuti v e readings, the interv al increases by 5 s per c ycle up to a ceiling of 60 s. When the units digit changes, indicating acti v e w ater o w , the interv al immediately drops to 5 s, then decreases by 2 s per c ycle do wn to a oor of 3 s. An anomaly e v ent immediately resets to the minimum interv al to enable rapid consecuti v e readings. This polic y cuts unnecessary inference c ycles during idle periods without sacricing responsi v eness. 3.8. MeterW atch inter net of things dashboard MeterW atch is a Flask web application that serv es as the monitoring front-end for the system. Figure 3 sho ws the complete hardw are–softw are architecture connecting the ESP32-CAM t o the MeterW atch dashboard. On each HTTP POST from the ESP32-CAM, the serv er recei v es the ra w JPEG image, the recognised meter v alue, mer ged-cell bounding box coordinates for display o v erlay , and the three rmw are ags ( confirmed , anomaly , sampling interval ). It dra ws bounding box o v erlays on the recei v ed image and stores all data in memory . Figure 3. System architecture connecting the ESP32-CAM, the Flask serv er , and the MeterW atch dashboard The dashboard frontend, refreshed via AJ AX polling e v ery 3 s , pro vides: a li v e image grid of the ten most recent unique readings with o v erlays. A Chart.js trend graph of the last 20 readings with anomaly mark ers highlighted in red. A connected de vice panel supporti ng multiple simultaneous ESP32-CAM units. Real-time session statistics (conrmed reading rate, anomaly rate, a v erage sampling interv al, system success rate). Also a full reading history table with per -ro w highlighting for anomalies, conrmed readings, and error states. Figure 4 sho ws t he MeterW atch web interf ace itself, illustrating the li v e reading grid, connected-de vice panel, session statistics, trend chart, and reading history table described abo v e. The de vice count, reading total and session statistics in this screenshot demonstrate interf ace functionality only and do not represent the single-meter e xperimental results reported in section 4. Int J Recongurable & Embedded Syst, V ol. 15, No. 3, No v ember 2026: 629–644 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Recongurable & Embedded Syst ISSN: 2089-4864 ❒ 637 Figure 4. MeterW atch dashboard interf ace (illustrati v e screenshot) 4. RESUL TS AND DISCUSSION 4.1. Experimental pr otocol All metrics were collected on a ph ysical Baylan-brand mechanical w ater me ter under the same real - w orld conditions used during data collection. The full 6,155-image dataset w as collected across approximately 6–7 distinct capture sessions. The test images were e xcluded from model training and INT8 calibration. Ho w- e v er , because the split w as performed at image le v el rather than acquisition-session le v el, the e v aluation should be interpreted as an internal frame-le v el v alidation rather than an independent session-le v el test. The train- test split w as performed by random image-le v el partitioning across the complete capture pool rather than by session-le v el separation. T est images may therefore share capture ses sions, lighting setups, and meter posi- tions with training images, dif fering at the frame le v el rather than representing fully independent acquisition conditions. This should be tak en into account when interpreting the reported accurac y . Ima g e pr ocessing and computer vision on low-cost micr ocontr oller s: a case study on ... (Sohr ab J alali) Evaluation Warning : The document was created with Spire.PDF for Python.
638 ❒ ISSN: 2089-4864 4.2. Classication perf ormance The INT8 model w as e v aluated on the 553-image box-annotated test subset, comprising 79,632 in- di vidually labelled grid cells (2,765 true digit cells and 76,867 true background cells) at a x ed classication condence threshold of 0.5. The model achie v es a 99.60% o v erall cell-le v el accurac y including the background class (79,310 of 79,632 ce lls correctly classied). Because background cells mak e up 96.5% of all e v aluated cells and are classied almost without error (99.69% recall), this gure is dominated by the background class and should not be read as a measure of digit-recognition dif culty . Restricting the e v aluation to the 2,765 true digit cells gi v es a 96.89% digit-only micro accurac y (2,679 of 2,765). Macro precision across the ten digit classes is 90.44%, macro recall is 96.92%, and macro F1 is 0.9355. Class-le v el recall ranges from 93.89% (digit “6”, the class most frequently confused with digits “5” and “7”) to 99.16% (digit “2”). Class-le v el F1 ranges from 0.9142 (digit “8”) to 0.9631 (digit “3”). These gures deri v e from a single Baylan-brand meter and an image-le v el split across approximately 6–7 capture ses sions. Multi-session and cross-brand v alidation are needed before treating this accurac y as a general performance bound. Complete v e-digit meter readings were correctly recognised in 485 of the 553 box-annotated test images, corresponding to an e xact-reading accurac y of 87.70%. Figure 5 sho ws the ro w-normalised confusion matrix on the 553-image test subset (79,632 grid cells) with per -class recall on the diagonal. The dominant confusion sources are misclassication of true-class sam- ples as background, reaching 3.45% for digit “8” and 2.62% for digit “6”, and genuine inter -digit confusion, most notably digit “6” being misread as digit “5” (1.75%) and digit “7” (1.31%), and digit “9” being misread as di git “8” (2.25%). F our digit-pair confusions e xceed 1% of the corresponding true class: “0” → “9” (1.17%), “6” → “5” (1.75%), “6” → “7” (1.31%), and “9” → “8” (2.25%). Background misclass ication is therefore not the sole error source. Figure 6 sho ws the per -class F1 distrib ution. Figure 5. Ro w-normalised confusion matrix of the INT8 model on the 553-image box-annotated test subset Int J Recongurable & Embedded Syst, V ol. 15, No. 3, No v ember 2026: 629–644 Evaluation Warning : The document was created with Spire.PDF for Python.