iTechGuides is reader-supported. When you buy through links on our site, we may earn an affiliate commission. As an Amazon Associate I earn from qualifying purchases. Learn more
PV electroluminescence (EL) models often miss small cracks because the crack pixels make up only a small part of an image, while background and larger cell features dominate training and evaluation. That imbalance can leave a model with high crack recall but poor crack boundaries: in a 2023 benchmark, DeepLabv3+ with custom class weights achieved 86% crack recall but only 25% crack mean intersection over union (mIoU). Fixing the problem requires checking crack-specific metrics and masks, addressing label and image-resolution issues, and validating changes on the acquisition conditions where the model will actually be used.
Why small cracks are hard to segment
Semantic segmentation assigns a class to each pixel. In an EL image, a crack may be a narrow line or a small branching structure surrounded by much larger areas of background, cell material, or other features. If the crack occupies very few pixels, a model can perform well on the abundant pixels while still producing an incomplete crack mask. Overall pixel accuracy can therefore look reassuring without showing whether the smallest defects were found.
The challenge is not only whether a model detects a crack somewhere. A useful segmentation must also trace its extent closely enough to match the label. Missing a short branch, breaking a thin line into disconnected pieces, or expanding a mask beyond the visible crack can all reduce overlap with the labeled region. For this reason, small-crack performance should be evaluated as a distinct problem rather than inferred from a whole-image score.
Windows Errors? Fix Them Before They Spread
Repair common Windows errors and clear accumulated junk for a smoother, more stable PC - no reinstall needed.Free scan · no reinstallOutdated Drivers Are Slowing You Down
One free scan finds every outdated or missing driver and matches the right update for your exact hardware.Free scan · exact hardware matchForeground-background imbalance
When crack pixels are scarce, a training objective dominated by the most common pixels may give the model too little incentive to learn the rare class. This is one reason class-aware loss weighting is worth testing. But changing the relative importance of crack pixels does not make the visual evidence clearer, resolve inconsistent labels, or guarantee an accurate boundary. Weighting is an experiment to evaluate, not a universal remedy.
#1 Best Overall
- 【After‑Sales Support & Service】: We offer 1-Year Manufacturer Support and 24-hour professional service. Contact our support team anytime if the device malfunctions due to product defects within 12 months after purchase.
- 【1920P HD Resolution】: Specialized for plumbers and mechanics, this 1920×1440 endoscope captures sharper images than standard 1080P. The ultra‑thin probe with adjustable LED lights inspects pipes, HVAC, and tight spaces effortlessly‑even in dark or damp conditions.
- 【Flexible Inspection Performance】: Optimized for multi‑angle observation, this endoscope helps you check pipe corners, wall wiring, and engine compartments. It delivers clear visuals for various maintenance and repair scenarios.
- 【Semi‑Rigid Cable Control】: The semi‑rigid cable bends smoothly yet maintains shape, letting you navigate curves in drain pipes or HVAC ducts without losing control during inspection work.
- 【Plug‑and‑Play Compatibility & Package Include】: Works with multiple smartphones. Includes detachable connectors for Type‑C, Micro USB and Lightning devices. Simply scan the QR code to download the app, then plug in to view real‑time inspection footage directly on your phone screen. Package includes: Endoscope Camera, Type‑C Port, Micro‑USB Port, Lightning Port, Magnet, Hook, Protective Cap, Instruction Manual.
The 2023 SEiPV-Net study likewise describes cracks as occupying a very small pixel range. It reports that weighting improved visual results, while micro-defect segmentation remained imprecise. This is an important distinction: an intervention can help without making the smallest targets reliably precise.
Small structures can be lost in the image pipeline
A narrow crack has little spatial width to spare. Resizing an image, cropping it into tiles, or changing the scale at inference can reduce a thin structure to fewer pixels or remove it at a crop boundary. These are practical failure modes to check when small-crack masks are weak. The studies summarized here do not establish one universally optimal resolution, crop size, or tile size, so compare such choices at the original evaluation scale rather than assuming a preferred setting.
Labels define what the model is asked to learn
Segmentation quality depends on what annotators mark as a crack. If one label includes a faint branch and another omits it, or if ambiguous edges are drawn differently, the model receives conflicting examples. Before changing the architecture, confirm the crack definition, how thin branches and uncertain boundaries are handled, and whether the labeling convention is applied consistently. An apparent model failure can be partly an annotation disagreement.
How to tell whether the problem is recall, overlap, or continuity
Use more than one crack-specific measure and inspect the masks. Recall indicates how much of the labeled crack the model finds; it does not by itself show whether the predicted pixels are tightly placed. IoU or Dice measures overlap between predicted and labeled regions, so it provides a different view of segmentation quality. Visual review helps expose errors that a single aggregate number hides, including disconnected lines and missed branches.
| Measure or check | What it helps answer | What it cannot establish alone |
|---|---|---|
| Crack recall | How many labeled crack pixels are detected. | Whether the predicted boundary is close to the label or whether extra pixels are being predicted. |
| Crack IoU or Dice | How much the predicted crack region overlaps the labeled region. | Whether performance is consistent across crack sizes, image sources, or acquisition conditions. |
| Visual mask review | Whether cracks are continuous, branches are present, and false detections are visible. | A reproducible summary across a large evaluation set unless paired with defined metrics and sampling. |
| Size-stratified results | Whether the smallest cracks perform worse than larger ones. | Why the difference occurs; diagnosis still requires reviewing images, labels, and preprocessing. |
The benchmark result illustrates why these views must be read together. Pratt and colleagues’ 2023 study in Systems and Soft Computing evaluated DeepLabv3+ with custom class weights and reported 86% crack recall alongside 25% crack mIoU. Recall alone would obscure the weak overlap score. The same paper found that larger feature classes had higher mIoU than smaller defect classes, consistent with the concern that aggregate performance can favor large regions.
Rank #2
- [1920P HD View for Tight Spaces] 1920P endoscope with a 7.9mm ultra-thin probe captures clear view at 1920×1440 resolution, helping you spot tiny cracks, leaks, and rust or damage deep inside pipes, engines, and walls. It's also a great leak detection tool.
- [Easy Plug-and-Play for iOS & Android Phones] Just download the app and connect the snake camera to your phone via Type-C, Micro-USB, or Lightning (with adapter). Real-time images appear on your screen in seconds—no pairing, no lag, and no complicated setup.
- [16.4ft Semi-Rigid Snake Cable & IP67 Waterproof Probe] The 16.4ft semi-rigid cable holds its shape so you can push the borescope around corners and down long pipes. The IP67 waterproof 7.9mm probe with 8 adjustable LED lights delivers bright, clear vision even in wet, dark environments.
- [Ideal Tool for Homeowners, Mechanics & Plumbers] Perfect for checking clogged drains, air conditioning ducts (HVAC), car engines, wall cavities, and other hard-to-reach areas around your home. This inspection camera helps you find problems quickly without unnecessary tearing things apart or paying for guesswork repairs.
- [Full Accessory Kit & Reliable Support Team] Package includes: endoscope camera with Type-C connector ×1, Micro-USB adapter ×1, Lightning adapter ×1, magnet ×1, hook ×1, protective cap ×1, side-view mirror ×1, and user manual ×1. If you have any questions, our support team is ready to help.
When reviewing individual masks, compare the image, reference label, and prediction side by side. Record whether the model misses a whole crack, loses only fine branches, truncates it at an edge, breaks it into segments, or predicts pixels outside the labeled path. These error types point to different checks: label consistency, spatial detail, class weighting, or false-positive control. Keep the review set fixed when comparing experiments so changes are not judged on different examples.
Why high recall can still mean a poor crack mask
Recall is about labeled crack pixels that the model recovered. It does not penalize extra predicted pixels in the same way that an overlap measure does. A prediction can capture much of the labeled line but also spread beyond its boundary, or find the main path while missing branches. Either pattern can produce an encouraging recall figure without a closely matched mask. The benchmark’s 86% recall and 25% mIoU make this difference concrete; they are not contradictory results, but measures of different aspects of the prediction.
Quick wins for a faster PC:
Clear out junk files and repair common Windows errorsFree Scan →Scan for outdated or missing drivers - takes under a minuteDriver Scan →When recall is high and IoU is low, inspect both omissions and additions rather than responding by simply increasing crack weighting. Check whether the label itself is narrow or inconsistent, whether predicted paths are too wide, and whether nearby features are being classified as crack. If recall is low too, look for entirely missed targets and review the smallest size groups separately. This diagnosis helps make the next experiment specific instead of changing several pipeline components at once.
A practical workflow for improving small-crack masks
- Write down the label rule. Define what counts as a crack, how thin branches are handled, and how uncertain boundaries are marked. Check a sample of labels from different annotators or batches for consistency before treating the model score as the only source of error.
- Build a crack-specific baseline. Evaluate on a held-out set and record crack recall plus IoU or Dice. Add visual review of false negatives and mask continuity, and report results by crack size so small targets do not disappear inside a class-wide average.
- Test class-aware loss or weighting. Give sparse crack pixels more influence as a controlled experiment. Compare the result against the baseline on the same split, watching precision and boundary quality as well as recall; a higher recall is not an improvement if it comes with an unacceptable burden of spurious crack pixels.
- Check spatial detail end to end. Review the image dimensions used for training and inference, resizing, cropping, and tiling. Confirm that narrow structures remain visible and that evaluation uses the intended original scale. Try alternate preprocessing choices as controlled comparisons; there is no single tile size or resolution established by the cited studies.
- Use unlabeled images if labels are the bottleneck. If the plant has plentiful unlabeled EL images but limited annotations, test a semi-supervised method against a supervised baseline using the same held-out evaluation split. Keep a representative labeled validation set for tuning rather than assuming a published gain will transfer.
- Test acquisition variation. Evaluate on images from distinct production batches or imaging conditions, and inspect errors by crack size, contrast, orientation, cell region, and image source. Preserve an untouched evaluation split for the final estimate instead of repeatedly tuning against it.
- Compare model families fairly. Use a common split, preprocessing, and metrics when comparing architectures. Report the protocol alongside results; otherwise a score difference may reflect data handling or labels rather than the model family.
Make weighting a controlled comparison
Start from a baseline with fixed labels, a fixed held-out split, and documented preprocessing. Then test a class-aware loss or custom weights while holding the rest of the setup steady. Compare crack recall, precision, IoU or Dice, and the visual quality of boundaries. Weighting can shift the balance between missed crack pixels and predictions outside the reference mask, so a rise in recall alone is not enough to choose a setting.
The evidence supports trying weighting, but it does not provide a universally successful weight value or guarantee improved micro-crack precision. The 2023 benchmark’s custom-weight DeepLabv3+ result and the SEiPV-Net study’s improved visual results are both accompanied by a caution: small defects can remain difficult. Select a setting using a representative validation set, then report its result on an untouched evaluation set.
Rank #3
- 2MP Global Shutter & 90fps High Frame Rate： This camera features a 2MP global shutter sensor and up to 90fps high-speed frame rate, effectively eliminating motion blur and distortion. It delivers stable and clear images for fast-moving objects, ideal for high-speed capture, motion detection and industrial applications.
- 2.8-12mm 4X Manual Varifocal Zoom Lens： Equipped with a 2.8‑12mm varifocal CS mount lens supporting 4X manual zoom. You can freely adjust focal length, focus and field of view to meet various needs from wide viewing to close‑up detail capture.
- Strong System & Device Compatibility： UVC compliant plug‑and‑play design with no driver required. Fully compatible with Windows, Linux, Jetson Nano and embedded systems, supporting stable long‑time working for industrial and daily use.
- Wide Software Support： Perfectly works with Lightburn, OpenCV, machine vision software, live streaming tools and video monitoring programs. Great for laser engraving monitoring, machine vision, production detection and live broadcast.
- Versatile Wide Applications： Widely used in industrial inspection, machine vision, Lightburn monitoring, USB video microscope, live streaming, high-speed recording, security monitoring and embedded projects.
Preserve detail without assuming one ideal crop
Check the full path from source EL image to prediction: original dimensions, any resizing, crop or tile construction, and inference scale. Compare the source image and processed input to ensure the smallest structures have not disappeared or become too faint to distinguish. A tile edge can also split a line, so review predictions near crop boundaries if tiling is used. These checks identify possible information loss; they do not imply that every deployment should use full-frame processing or a particular tile dimension. The cited publications do not name one best resolution or tiling recipe.
What published results do—and do not—show
The available studies offer useful evidence for choosing experiments, but they use specific datasets and protocols. Their percentages are not forecasts for another module type, camera, plant, or annotation scheme.
| Study or resource | Reported scope and result | How to interpret it |
|---|---|---|
| Pratt et al., Systems and Soft Computing, 2023 | 593 labeled cell images with pixel-level labels for 24 defects and features; 50 test images. DeepLabv3+ with custom class weights reported 25% crack mIoU and 86% crack recall. | A benchmark example of recall and overlap telling different stories. The paper also reports stronger mIoU for larger feature classes than for smaller defect classes. |
| SEiPV-Net, Energies, 2023 | Reports improved visual results with weighting, while micro-defect segmentation remained imprecise. | Supports testing weighting, but not treating it as a complete fix for fine defects. |
| PV-S3, Engineering Applications of Artificial Intelligence, 2025 | On UCF-EL, using 20% labeled samples, the semi-supervised study reports absolute improvements over its supervised comparison of 9.7 percentage points in mIoU, 13.5 in precision, 29.15 in recall, and 20.42 in F1. | These are gains reported for that study’s dataset and protocol. They should not be presented as expected gains on a different camera, plant, or split. |
| MultiSolSegment, Solar Energy, 2026 | Reports a multi-channel U-Net approach for overlapping EL-image features and a dataset of 2,340 labeled monocrystalline-silicon images. | A newer direction relevant when image features overlap; the reported work does not establish that this architecture is universally best for small cracks. |
| MSPS-Dataset repository, accessed 2026 | Describes 1,500 production-line EL crops at 640 × 590 pixels, with 1,000 training images and 500 test images. The repository identifies minute targets, low contrast, scale variation, irregular boundaries, and severe foreground-background imbalance as challenges. | Useful repository-described dataset and acquisition facts, not a peer-reviewed estimate of deployment performance. |
The PV-S3 result makes semi-supervised learning worth considering when annotation is scarce, but its 20% labeled-sample setting and reported gains belong to its UCF-EL experiment. Reproducing a percentage elsewhere would require matching the relevant data and protocol; a different image source may produce a different outcome.
MultiSolSegment is also relevant if the task includes overlapping EL-image features. Its multi-channel approach and reported monocrystalline-silicon dataset broaden the model directions available to compare, but the publication detail alone does not settle small-crack performance for another dataset. Treat it as a candidate for a controlled comparison, not a reason to skip local evaluation.
When annotations are limited
More labeled images are not always immediately available, particularly when pixel-level masks require careful marking of thin branches and ambiguous edges. If unlabeled EL images are plentiful, semi-supervised learning is a candidate experiment rather than a substitute for checking labels. In a 2025 study in Engineering Applications of Artificial Intelligence, PV-S3 used 20% labeled samples on UCF-EL and reported absolute gains over its supervised comparison of 9.7 percentage points in mIoU, 13.5 in precision, 29.15 in recall, and 20.42 in F1.
The Tool Desk
Outbyte Driver Updater FREEScan for outdated or missing drivers - takes under a minuteDriver Scan →Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →Those figures describe that dataset and protocol, not a guaranteed return from adding unlabeled data. A local test still needs a representative labeled validation set and the same held-out evaluation protocol used for the supervised comparison. If the validation images do not cover the small cracks, contrast levels, and acquisition sources that matter, a score improvement may not answer the deployment question.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.How to make evaluation reflect deployment
A random held-out set is useful only if it represents the conditions where predictions will be used. EL image resolution, contrast, defect scale, module type, and labeling conventions can vary across sources. The MSPS repository specifically describes production-line data with low-contrast, multiscale, irregular defects and foreground-background imbalance, illustrating why a model can face several simultaneous challenges.
- Separate tuning from final evaluation. Keep an untouched split for the final result; use a separate validation set to choose weights, preprocessing, and model settings.
- Record the data protocol. State how images were split, what preprocessing was applied, and which metrics were calculated. This makes comparisons interpretable and helps prevent data handling differences from being mistaken for model improvements.
- Break down performance by target size and image source. Include the smallest cracks and distinct batches in the report rather than relying on an overall average.
- Inspect false positives as well as misses. A model that labels many non-crack pixels can appear more complete while making the mask less useful. Review both the missed labeled regions and predicted regions outside the reference mask.
- Recheck the annotation convention across domains. If images from a new source are labeled differently, a score change may reflect the target definition as well as the model’s ability to generalize.
Compare models on more than an overall score
A fair architecture comparison uses the same dataset split, preprocessing, label convention, and metrics. Alongside crack recall and IoU or Dice, compare continuity, performance in the smallest crack-size groups, false-positive burden, and stability across imaging batches or camera domains. Include annotation needs and inference cost if they matter to deployment; the cited studies do not give a universal cost figure or show that one architecture is always cheapest.
MultiSolSegment, published in Solar Energy in 2026, reports a multi-channel U-Net for overlapping EL-image features and 2,340 labeled monocrystalline-silicon images. That makes it a relevant model family to consider when features overlap, not proof that it will outperform another model on a particular crack dataset. A comparison is useful only when the local data and evaluation protocol are held constant.
Do these 3 things before closing this tab:
1Repair Windows errors before they cause bigger problems2Scan for outdated or missing drivers - takes under a minute3Clear out junk files and repair common Windows errorsChoosing the next experiment
Use the failure pattern to choose a small, interpretable experiment rather than changing several parts of the pipeline at once. If recall is low and visual review shows entire small cracks missing, start by checking labels, class imbalance, and whether preprocessing preserves the structures. If recall is high but IoU is poor, examine boundary placement and extra predicted pixels before increasing crack weighting further. If the masks look acceptable on one batch but degrade on another, compare image conditions and labels by source, then validate on representative data from those conditions.
Best Value
For architecture comparisons, keep the dataset split and evaluation protocol fixed. Compare crack recall, IoU or Dice, continuity, size-stratified results, false-positive burden, and robustness across batches. Also track annotation needs and inference cost if those affect deployment; the published figures summarized above do not establish a universal winner on either measure. The purpose of a comparison is to discover what works under the local conditions, not to transfer a leaderboard result without checking its assumptions.
A strong small-crack segmentation report therefore includes more than a headline score: the label definition, held-out protocol, crack-specific overlap and recall, size breakdown, visual error review, and results across relevant acquisition conditions. Those details expose whether an apparent improvement actually finds fine cracks and traces them accurately.
Frequently Asked Questions
Why does my solar panel crack detection model miss small cracks?
In EL images, crack pixels can be sparse compared with background and larger cell features, so aggregate performance can hide weak results on small targets. Resizing or cropping may also remove narrow structures, and inconsistent annotations can give the model conflicting examples. Check crack-specific recall and overlap, inspect missed masks, and report results by crack size.
Why is crack recall high but IoU low?
Recall measures how much of the labeled crack was found; it does not alone show how tightly the prediction follows the label or how many extra pixels were predicted. In Pratt et al.’s 2023 benchmark, DeepLabv3+ with custom class weights had 86% crack recall but 25% crack mIoU. Review both missing branches and pixels outside the reference mask.
How can I handle class imbalance in PV cell defect segmentation?
Test class-aware loss or weighting as a controlled comparison, then measure crack recall, precision, IoU or Dice, and boundary quality on the same split. Weighting improved visual results in the 2023 SEiPV-Net study, but micro-defect segmentation remained imprecise, so it is not a guaranteed fix.
Can semi-supervised learning help when I have few crack labels?
It is worth evaluating if unlabeled EL images are plentiful. The 2025 PV-S3 study reported gains on UCF-EL using 20% labeled samples, but those results are specific to its data and protocol. Compare against a supervised baseline on the same held-out split and keep a representative labeled validation set.
Is there a best image resolution, tile size, or model architecture for small cracks?
The cited studies do not establish a universally best resolution, tile size, or architecture. Preserve spatial detail, compare preprocessing choices at the intended evaluation scale, and assess candidate models on a common split with crack-specific metrics and size-stratified results.
The Bottom Line
Small-crack segmentation is a sparse-target problem, so do not trust an overall score or recall alone. Define consistent labels, inspect predictions, report crack-specific overlap and recall by size, test weighting and preprocessing as controlled experiments, and validate across the image conditions where the model will be used.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
