💻 computer science

STC-Net: Electroluminescence-Based Solar Cell Crack Segmentation for Power Loss Estimation

This paper introduces STC-Net, a novel deep learning framework that leverages edge and spectral priors along with a boundary-topology refinement module to accurately segment thin, elongated cracks in electroluminescence images, thereby enabling reliable power loss estimation for photovoltaic cells.

Original authors: Shanaka Ramesh Gunasekara, Akila Eranda Devanarayana, Imasha Guruge, Nuwantha Fernando, Ehsan Asadi

Published 2026-08-04

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Shanaka Ramesh Gunasekara, Akila Eranda Devanarayana, Imasha Guruge, Nuwantha Fernando, Ehsan Asadi

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the sun as a giant, free battery charger for the world. We've built millions of solar panels to catch that light and turn it into electricity for our homes and gadgets. But just like a human body, these panels can get sick. They can develop tiny, invisible cracks that don't look like much to the naked eye but can stop the flow of power, much like a kink in a garden hose stops the water. To find these hidden injuries, scientists use a special kind of "X-ray vision" called Electroluminescence (EL). Instead of taking a photo in the dark, they make the solar cells glow like little lightbulbs. If a cell is healthy, it shines brightly and evenly. If it has a crack, that part goes dark, revealing a ghostly, jagged line where the electricity is leaking away.

The big challenge is that these cracks are incredibly tricky to spot. They aren't big, round blobs like a bruise; they are thin, long, and often broken into pieces, like a spiderweb or a lightning bolt frozen in time. Old computer programs trying to find them often got confused, either missing the tiny lines entirely or drawing them all wrong. This matters because if we don't know exactly where the cracks are, we can't tell how much power the panel is actually losing. It's like trying to fix a leaky roof without knowing exactly where the water is dripping.

Enter STC-Net, a new computer brain designed by researchers to solve this specific puzzle. Think of STC-Net as a super-smart detective who doesn't just look at the picture but also studies the "edges" of the lines and the "texture" of the light. While other programs might just guess where a crack is, STC-Net uses three special tools to get it right. First, it has an "Edge Eye" that is obsessed with sharp boundaries, ensuring it doesn't miss the thin, hairline fractures. Second, it has a "Spectral Ear" that listens for the faint, high-pitched signals of broken patterns that other tools ignore. Finally, it has a "Topology Brain" that understands that a crack is a connected path, not just a random dot, helping it stitch broken pieces back together into a complete line.

The researchers tested this new detective on a dataset of over 1,200 solar cell images. The results were impressive. During its training, STC-Net achieved a 95.98% accuracy in matching the cracks exactly where they were supposed to be (a score called MIoU) and a 98.01% score in how well it overlapped with the real cracks (MDice). Even when it faced brand-new, unseen images for the first time, it still managed a 72.52% accuracy in location and 80.16% in overlap, proving it didn't just memorize the answers but actually learned the rules.

But STC-Net doesn't stop at just drawing the cracks; it also acts as a fortune teller for the panel's health. Once it maps out the cracks, it calculates how much "dead space" those cracks create. It then translates that dead space into a power loss estimate. For example, if the model finds a crack that covers a tiny 0.250% of the cell, it estimates a power loss of just 0.0025 Watts. However, if it finds a massive, degraded area covering 41.3% of the cell, the estimated power loss jumps to 0.413 Watts. This creates a direct link between a picture of a crack and a real number representing lost energy.

The paper suggests that this approach is a significant step forward because it moves beyond just saying "there is a crack" to actually quantifying "how much power is gone because of this crack." While the current model is a simplified estimate and doesn't yet account for every complex electrical factor in a real-world power plant, it successfully bridges the gap between taking a photo of a broken cell and understanding its impact on the energy grid. By combining sharp edge detection with smart topology, STC-Net offers a practical way to keep our solar batteries healthy, ensuring they keep charging our world for years to come.

Technical Summary: STC-Net for Solar Cell Crack Segmentation and Power Loss Estimation

Problem Statement

Accurate assessment of cracks in photovoltaic (PV) modules is critical for reliability analysis and predictive maintenance. While Electroluminescence (EL) imaging provides high-contrast visualization of structural defects, existing deep learning segmentation methods often struggle to capture the specific characteristics of solar cell cracks. Unlike generic objects in natural images, cracks are thin, elongated, structurally constrained, and often fragmented yet connected. Conventional attention-based or general-purpose segmentation architectures (e.g., U-Net, DeepLabv3+) frequently fail to preserve crack continuity and boundary precision due to the severe imbalance between crack pixels and background, as well as the lack of explicit modeling for crack topology. Furthermore, most existing EL-based studies focus on defect classification or coarse severity assessment, failing to explicitly link pixel-level defect localization to practical electrical performance degradation, such as power loss estimation.

Methodology

The authors propose STC-Net (Solar Topology Crack Network), a lightweight, end-to-end framework designed to address these limitations through a multi-branch encoder-decoder architecture enhanced by specific priors and a refinement module.

1. Architecture Overview

STC-Net follows an encoder-decoder structure with two auxiliary prior extraction branches and a dedicated prediction refinement module:

Encoder: A hierarchical U-Net-like encoder extracts multi-scale semantic representations from the input EL image.
Auxiliary Prior Branches:
- Edge Prior: Utilizes fixed Sobel filters to compute gradient magnitude maps, explicitly preserving boundary discontinuities and thin line-like structures. These are processed through an edge pyramid to provide multi-scale boundary cues.
- Spectral Prior: Constructs a spectral residual map using Fourier transforms and high-pass filtering to suppress low-frequency background components and emphasize faint, irregular crack patterns. These features are also processed through a spectral pyramid.
Topology-Guided Decoder: The decoder reconstructs high-resolution features by fusing semantic encoder features with the edge and spectral priors. A lightweight attention unit dynamically predicts fusion weights to adaptively combine these sources, ensuring the decoder is guided by semantic context, geometric boundaries, and high-frequency structural cues.
Prediction Refinement Module: Instead of a single output, the network generates three intermediate logits via a coarse head, an edge head, and a topology head. These are fused by a boundary-topology refiner to produce a refined representation before the final segmentation head generates the pixel-wise mask.

2. Learning Objective

The network employs deep structural supervision. The total loss function combines:

Segmentation Loss: A weighted sum of Dice loss and Binary Cross-Entropy (BCE) for the final mask.
Auxiliary Loss: Additional supervision applied to the intermediate coarse, edge, and topology heads to encourage the learning of accurate regions, boundaries, and topologically consistent structures.

3. Power Loss Estimation

To bridge segmentation with performance assessment, the framework includes a module to estimate power loss:

Inactive Area Proxy: The predicted crack mask is dilated to form an influence zone. This is combined with a "dark-region" mask (derived from low-intensity pixels in the EL image) to estimate the electrically inactive area associated with the crack.
Power Calculation: The ratio of the inactive area to the total cell area is calculated. Assuming a linear relationship, the power loss ( $\Delta P$ ) is estimated as a fraction of the nominal power ( $P_{nom}$ ) proportional to this inactive area percentage.

Key Contributions

Specialized Framework: Introduction of a lightweight, end-to-end framework specifically designed to capture the geometric and topological nature of crack defects in EL images, maintaining low computational complexity.
Topological and Orientation Awareness: Integration of orientation-aware and topology-guided modules (edge and spectral priors, multi-head refinement) to enhance continuity preservation and detect thin, elongated, and directionally varying crack patterns.
Performance-Aware Assessment: Extension of the framework from visual segmentation to practical condition evaluation by estimating power loss directly from predicted defect regions, linking image-based detection to electrical degradation.

Experimental Results

The model was evaluated on the PVEL-S dataset, a refined subset of 1,200 polycrystalline silicon cell images (840 training, 360 test).

Segmentation Performance:
- Training: STC-Net achieved 95.98% MIoU, 98.01% MDice, and 98.00% MAcc. While CAAK-Net reported a slightly higher MIoU (96.07) compared to STC-Net (95.98), STC-Net achieved the highest MDice (98.01) and MAcc (98.00) among the compared methods.
- Inference (Unseen Test Data): Achieved 72.52% MIoU and 80.16% MDice, demonstrating robust generalization.
- Efficiency: The model operates at 41.38 FPS, indicating computational practicality.
Qualitative Analysis: Visual results show that STC-Net effectively preserves the morphology of thin cracks and large degraded regions. For fine cracks, the model achieved IoU values of 0.816–0.843, while larger regions showed IoU values between 0.938 and 0.985.
Power Loss Estimation: The framework successfully correlated crack severity with estimated power loss. For example, samples with low inactive areas (~~0.25%) showed negligible power loss (~~0.0025 W), while high-severity samples (up to 41.3% inactive area) showed corresponding higher losses (~0.413 W).

Significance and Claims

The paper claims that STC-Net provides a significant advancement by explicitly modeling the structural characteristics of cracks rather than treating them as generic regions. The primary significance lies in the dual capability of the framework:

Accurate Localization: It offers superior boundary preservation and continuity for thin, fragmented defects compared to existing general-purpose segmentation models, achieving top-tier performance in MDice and MAcc metrics.
Practical Utility: It establishes a direct, interpretable link between visual defect segmentation and PV degradation assessment. By estimating power loss from the segmented inactive area, the framework supports inspection, maintenance prioritization, and reliability monitoring.

The authors note that the power loss estimation is a "simplified surrogate" and does not capture complex nonlinear electrical behaviors (e.g., local heating or interconnection effects). However, it serves as a practical first-order bridge between image analysis and engineering-oriented performance evaluation. Future work is proposed to validate the framework against measured electrical data and extend the analysis to module-level power loss.