2.1 Traditional image processing-based Mura defect detection methods

Early traditional defect detection methods primarily relied on thresholding, edge information, feature extractors, and classifiers for defect classification and localization. Sun et al. [18] proposed a cascaded Mura detection approach based on mean shift and level set algorithms, where the mean shift algorithm is first used to locate potential defect regions, followed by segmentation using the level set method. Chen et al. [19] addressed the challenge of detecting Mura defects with low brightness contrast against the background by proposing a background reconstruction method based on discrete cosine transform (DCT), generating a background image separated from the defect. They further quantified the saliency of Mura defects using the Just Noticeable Difference (JND) model. Tsai et al. [20] tackled the difficulties of detecting defects in large-size, high-density LCDs with complex and irregular patterns by applying Fourier-based image reconstruction to remove periodic backgrounds from 1D line patterns, enabling accurate segmentation of local anomalies. Cui et al. [21] employed three visual-based techniques for display defect detection and used OTSU thresholding to segment multiple defect types across various uniform background colors. Chen et al. [22] proposed a salt-and-pepper defect detection method using mean filtering and statistical control charts. The method first identifies defect pixels via mean filtering and binarization, then monitors the number of detected pixels with a control chart to determine panel-level defects. However, these traditional methods typically rely on background regularity, involve high computational costs, and have limited adaptability in complex or variable scenarios. Moreover, their accuracy is often unstable, making them unsuitable for direct deployment on industrial production lines.

2.2 Deep learning-based Mura defect detection methods

With the rapid development of deep learning, many methods have adopted end-to-end training frameworks that enable efficient extraction of display defect features and semantic segmentation. In the field of LCD defect detection, deep learning-based approaches have made notable progress, with research primarily centered on the optimization and application of general object detection models. Zhu et al. [23] employed YOLOv3 as the detection framework to identify defects and noise under varying backgrounds and viewing angles. By subtracting the noise detection results from the defect results, the method achieved a more accurate defect localization. Çelik et al. [24] analyzed deep learning-based LCD defect detectors using RetinaNet, YOLOv3, and M2Det, and found that RetinaNet offered the best balance between accuracy and processing time. Luo et al. [9] improved the YOLOv5 backbone by widening both shallow and deep convolutional layers, and proposed a shallow semantic fusion module to enhance the utilization of low-level features. They also introduced a contrast-enhancing attention mechanism to strengthen feature representations across spatial and channel dimensions.

Furthermore, extensive efforts have been made to design efficient neural network architectures that reduce computational overhead while improving detection accuracy. Lin et al. [25] used MobileNet as a baseline and incorporated a channel attention mechanism to enhance channel-wise features for Mura defect detection in LCDs. Chen et al. [14], building upon the YOLOv4-tiny framework, integrated atrous spatial pyramid pooling and depthwise separable convolutions to expand the receptive field, while spatial and channel attention modules were applied to further enhance feature maps. These improvements enable faster and more precise defect detection, supporting real-time applications in resource-constrained environments.

3.1 Overview

Figure 2 Overall architecture of the proposed SIFNet. It consists of six key components: Mv2 encoder, Graph-based Semantic Interscale-fusion Block (GSIB), Semantic Fluid Aggregation Module (SFAM), Semantic Graph Inference Module (SGIM), NeighborFusion Edge Enhancement Module (NEEM), and lightweight decoder.

As shown in Figure 2, the proposed lightweight SIFNet is built upon a common encoder-decoder architecture [29, 30] and consists of six key components: Mv2 encoder, Graph-based Semantic Interscale-fusion Block (GSIB), Semantic Fluid Aggregation Module (SFAM), Semantic Graph Inference Module (SGIM), NeighborFusion Edge Enhancement Module (NEEM), and lightweight decoder. For the encoder of SIFNet, we adopt the lightweight MobileNet-V2 [17]. Specifically, we retain the first 17 inverted residual bottlenecks while truncating the last two convolutional layers and average pooling layer to make it suitable for Mura detects detection. We divide MobileNet-V2 into five blocks based on the first, third, sixth, 13th, and last bottlenecks. The output five-level features are denoted as $F_i\in R^{c_i\times h_i\times w_i}$, $i$ = 1,2…5, where $h_i$ and $w_i$ are $\left(640/2^i\right)$, and $c_i\in \{16,24,32,96,320\}$. Next, to fuse multi high-level semantic features and model abstract semantic concepts for more precise defect localization, the features $\{F_3,F_4,F_5\}$ are processed by the GSIB, which integrates SFAM and SGIM to extract global contextual information and aggregate cross-scale semantics via hierarchical flow fusion and graph convolution. In addition, NEEM explores the correlation between shallow features $\{F_1,F_2\}$, enhancing the model's ability to perceive fine-grained structures and represent spatial details more effectively. Furthermore, edge information is extracted using a pooling subtraction operation , and the Canny operator is applied to generate edge supervision for correcting edge errors. Finally, three decoders are employed in a progressive manner to emphasize target objects and restore resolution.

3.2 Graph-based Semantic Interscale-fusion Block

Mura defects typically exhibit large-area distribution, accompanied by uneven brightness variations and blurred boundaries, making precise segmentation challenging. To address these issues, we propose the Graph-based Semantic Interscale Fusion Block (GSIB), which integrates high-level semantic information across spatial scales and efficiently extracts semantic cues from the spatial context. By combining graph-based modeling with a soft attention mechanism, GSIB excels at capturing long-range dependencies and maintaining global semantic consistency. It effectively activates high-level features in potential defect regions and enhances synergy between spatial location and semantic representation, enabling accurate modeling of global feature relationships within Mura-affected areas. GSIB improves the model's ability to perceive large-scale brightness gradients while maintaining low computational complexity and parameter overhead, achieving a strong balance between performance and efficiency. As shown in Figure 2, GSIB comprises two key components: the Semantic Fluid Aggregation Module (SFAM) and the Semantic Graph Inference Module (SGIM).

(1) Semantic Fluid Aggregation Module. SFAM employs a holistic strategy to capture global semantic features, enabling comprehensive extraction of semantic cues across the entire image. As shown in Figure 2, taking the encoder feature $F_5\in {ℝ}^{C\times H\times W}$ as an example, SFAM introduces a global attention mechanism to selectively extract visual primitives along the spatial dimension, thereby generating a semantic flow representation $F$. Specifically, $F_5$ is passed through two parallel $1\times 1$ convolutional branches to produce a feature query $F_{\text{q}}\in {ℝ}^{C\times HW}$ and attention weights $F_{\omega }\in {ℝ}^{HW\times N}$. The attention branch applies a softmax function to normalize the spatial dimension, resulting in a saliency distribution across the spatial domain. Finally, matrix multiplication is used to integrate the attention-guided global semantic representation, producing the semantic flow feature $F\in {ℝ}^{C\times N}$:

where $\otimes$ denotes matrix multiplication. $\phi$ and $\theta$ represent two independent $1\times 1$ convolution operations, with $W_{\phi }$ and $W_{\theta }$ denoting their respective convolution parameters.

Due to the semantic gap between high-level features at different scales, direct multi-scale fusion may amplify inconsistencies between adjacent layers. To address this issue, the semantic flow $F$ is further refined using a soft attention mechanism that selectively emphasizes informative input features before passing them to the SGIM for interscale fusion. Specifically, the input feature $F_5$ is upsampled and processed through a $1\times 1$ convolution, followed by a softmax operation along the channel dimension, producing an $N$-dimensional attention map $F_{\text{att}}\in {ℝ}^{N\times kH\times kW}$, where $k=2$ in our experiments. Guided by this attention map, the semantic flow $F\in {ℝ}^{C\times N}$ is adaptively integrated into the spatial domain, resulting in a semantic description map $F_s\in {ℝ}^{C\times kH\times kW}$. Finally, the semantic description map $F_s$ is fused with the original high-level feature $F_5$ via element-wise addition to obtain the final output of SFAM, denoted as $F_{sfam}\in {ℝ}^{C\times kH\times kW}$. The computation is formally defined as:

where $UP(\cdot )$ denotes the upsampling operation, and $\varphi$ represents the $1\times 1$ convolution.

In summary, SFAM effectively enhances global context understanding by integrating semantic flow with high-level features, while the use of soft attention helps bridge the semantic gap between features at different scales, providing a solid foundation for subsequent feature fusion in SGIM.

(2) Semantic graph-inference module. When dealing with complex Mura defect regions, traditional feature fusion methods often struggle to effectively capture the diversity of brightness variations and the inherent structural ambiguity. Therefore, we propose the Semantic Graph Inference Module (SGIM), which introduces an efficient graph-based feature aggregation mechanism to integrate receptive field information across multiple feature levels. Specifically, SGIM constructs high-order interscale relationships within the graph convolution domain, enabling explicit modeling of semantic dependencies through graph convolution operations. For Mura defects with smooth brightness transitions and blurred boundaries, SGIM effectively captures global brightness variation patterns across spatial regions, significantly improving both detection accuracy and robustness. Additionally, SGIM employs depthwise separable convolution blocks to reduce computational overhead. The detailed architecture is illustrated in Figure 2.

Taking the example shown in Figure 2, the SFAM output feature $F_{sfam}$ and encoder feature $F_3$ are first processed through two learnable linear mapping functions, ${\phi }_1$ and ${\phi }_2$, producing spatially aligned feature sequences $S_1$ and $S_2$, respectively: $S_1={\phi }_1\left(F_{sfam}\right),S_2={\phi }_2\left(F_3\right)$.

Next, a softmax function is then applied to $S_1$ to generate the attention map $W$: $W=\text{Softmax}\left(S_1\right)$.

The attention map is then multiplied with $S_2$, and the resulting features are fed into a Graph Convolutional Network (GCN) [39] to learn high-order semantic relationships among regions (sets of pixels with similar features). To reconstruct the graph domain features back into their original structural features, the inner product of GCN($\cdot$) is computed and transformed back into a 2D image feature map with the same dimensions as the original features using a linear mapping function ${\varphi }_f$. Finally, the reconstructed features are then combined with feature $F_3$ to produce the final output of SGIM, $F_{sgim}$, which is then passed to SFAM (if present) and the decoder. Notably, the number of nodes in the GCN is set to 16.

3.3 NeighborFusion Edge Enhancement Module

Shallow features contain rich texture details and edge information, which are critical for accurately delineating Mura defect boundaries, but they also introduce considerable background noise. Moreover, Mura defects typically exhibit gradual brightness transitions characterized by smooth intensity variations and indistinct boundaries without clear geometric structures. These features pose significant challenges for most existing defect detection methods. Since Mura defects typically exhibit gradual brightness transitions, their smooth intensity changes and blurred boundary characteristics often lack discernible geometric structures, making such regions particularly challenging for most methods to handle effectively. Thus, we propose the NeighborFusion Edge Enhancement Module (NEEM), which is designed to capture spatial correlations across multilevel features and enhance the model's sensitivity to boundary features in regions with uneven brightness. This enables more accurate identification of Mura defects in complex display environments. Additionally, NEEM is lightweight and computationally efficient, providing a practical solution for large-area Mura defect detection. The specific implementation details are illustrated in Figure 2.

First, the encoder features $F_2$ and $F_1$ are spatially aligned through upsampling. Depthwise separable convolutions (DSConv) are then applied to adjust their channel dimensions to match those of $F_2$, with spatial attention independently applied to each feature to enhance detail representation and prepare for subsequent fusion. The resulting feature maps are denoted as $F_1^{\prime }$ and $F_2^{\prime }\in {ℝ}^{24\times 320\times 320}$.

To facilitate information interaction between different feature levels, CrossNorm [40] is employed to compute and exchange the mean and variance of $F_1^{\prime }$ and $F_2^{\prime }$, enabling structured context transfer and improving the perception of fine-grained structures. Finally, the two refined features are fused via concatenation, and a residual connection is introduced to generate the final enhanced representation, improving the consistency of edge depiction and semantic expression. This process can be expressed as follows:

where $cat(\cdot )$ indicates channel-wise concatenation.

Next, a boundary enhancement module is applied to further refine the representation of edge features in Mura defect regions, generating the NEEM output $F_{neem}$, which is subsequently passed to the decoder. This process is defined as follows:

where $\text{Sigmoid}(\cdot )$ denotes the activation function, and $\oplus$ represents element-wise addition.

3.4 Decoder and Loss Function

Mura defects often exhibit diverse scales and brightness variations across different regions of a display, necessitating precise feature fusion at multiple levels. Therefore, we introduce a lightweight multi-scale decoder that adaptively integrates hierarchical features while progressively restoring spatial resolution. This approach enables accurate detection of Mura defects across varying scales, ensuring reliable localization and classification of both small, subtle imperfections and large-area brightness non-uniformities. Each decoder follows a consistent architecture, as illustrated in Figure 2. Specifically, each decoder sequentially consists of a DSConv layer, a Concate layer (if present, to merge NEEM or SGIM outputs), another DSConv layer, an upsampling layer, a final DSConv layer, and a Head layer. The final DSConv layer provides the decoder's output for the next stage, while the Head layer generates the prediction output for deep supervision. Specifically, the Head layer consists of a dropout layer and a $1\times 1$ convolutional layer, which generate three saliency maps with different resolutions: $f_3^{pre}\in {[0,1]}^{1\times 80\times 80}$, $f_2^{pre}\in {[0,1]}^{1\times 320\times 320}$, and $f_1^{pre}\in {[0,1]}^{1\times 640\times 640}$. The first two are used for deep supervision, while the last serves as the final output of SIFNet.

After obtaining the saliency maps $\left\{f_i^{pre}\right\},i=1,2,3$ and the edge prediction map $\left\{f_e^{pre}\right\}$, we train our SIFNet using a combination of BCE loss and IoU loss [33]. Therefore, our total loss consists of two parts, that is, the saliency loss and the edge loss. Moreover, we introduce a loss weight to balance the two losses, enabling more effective training. The total loss function $L_{total}$ can be formulated as follows:

where $L_{bce}$ and $L_{iou}$ represent the BCE loss and IoU loss, respectively, which are used to supervise $f_i^{pre}$ with the ground truth $G$. $G_e$ denotes the edge supervision map, generated from $G$ using the Canny operator.

4.1 Experiment Settings

Implementation Details. We use MobileNetV2 as the encoder, pre-trained on the ImageNet dataset , and the other modules are randomly initialized. For training and testing, all input images are resized to 640$\times$640, and random flipping and rotation are applied to augment the training data. We use Adam as the optimizer, with the learning rate initialized to 1e-4 and then scaled down by 10 every 30 epochs. Our model is trained end-to-end using PyTorch on an NVIDIA A100-SXM for 100 epochs with a batch size of 8.

Datasets. Given the limited availability of Mura defect samples from display screens, we have constructed three specialized datasets for Mura defect detection: the black screen scenario (Uneven_Black), the white screen scenario (Uneven_White), and a mixed black-and-white screen scenario (Uneven_Mix). All images have a resolution of 3612$\times$1358 pixels. The Uneven_Black dataset consists of 350 images exhibiting uneven displays on a black background, with 245 images allocated for training and 105 for testing. The Uneven_White dataset also comprises 350 images captured on a white background, following the same training/testing split. The Uneven_Mix dataset combines scenes from both black and white backgrounds, comprising 700 images in total, with 490 used for training and 210 for testing. Representative image samples are illustrated in Figure 3.

Figure 3 Examples of display defect under black and white screen conditions. (a) the black background scenario, (b) the white background scenario. The images from left to right are the original image, the contrast-enhanced image, and the corresponding ground truth mask, respectively.

Evaluation Metrics. To comprehensively evaluate the performance of the model, we performed an extensive evaluation from both the object-level and pixel-level perspectives. The details of these evaluation metrics are as follows:

(1) Object-level:

Precision and Recall. In object detection, precision refers to the proportion of correctly predicted positive samples among all samples predicted as positive, while recall represents the proportion of correctly predicted positive samples among all actual positive samples. They are defined as follows:

where TP (True Positive) refers to correctly detected objects, FP (False Positive) denotes incorrectly identified objects, and FN (False Negative) indicates actual objects that were missed by the model.

(2) Pixel-level:

Precision and Recall. In pixel-level detection, precision is defined as the ratio of correctly predicted regions to the total predicted regions, while recall is defined as the ratio of correctly predicted regions to the total ground truth regions. The formal definitions are as follows:

where $S\left(T\right)$ represents the predicted object region by the model and $G$ denotes the ground truth object region.

S-measure[35] is used to evaluate region-aware($S_r$) and object-aware($S_o$) structural similarity between predictions and GT and defined as:

where $\alpha$ is set to 0.5.

F-measure[36] is a holistic metric that considers both precision (P) and recall (R), which is defined as:

where $\beta$ is the balance parameter and ${\beta }^2$ is set as 0.3.

E-measure[37] is used to measure pixel-level matching and image-level statistics, which is defined as:

where $\varphi FM$ denotes the enhanced-alignment matrix and N is the total pixels of the image.

Mean absolute error[38] is to calculate the average absolute error of the prediction of salient objects (P) and ground truth (G), which is defined as:

4.2 Comparison with State-of-the-Art Methods

We conducted a systematic comparative analysis of recent visual models that have shown strong performance in object detection and instance segmentation tasks. In this study, we selected five representative versions from the YOLO series: YOLOv7 [26], YOLOv8 [16], YOLOv9 [27], YOLOv10 [28], and YOLOv11 [15] as our research targets. These models are built on advanced deep learning architectures and significantly improve detection accuracy and segmentation quality while maintaining high real-time performance. They have been widely applied in industrial settings and defect detection tasks, especially in scenarios that require both speed and precision, such as display panel inspection and surface quality analysis. We carried out a comprehensive evaluation of these models across several aspects, including object-level and pixel-level performance, computational efficiency, and generalization ability.

Quantitative Results. Tables 1, 2, and 3 summarize the quantitative results of our proposed method and competing methods on the three challenging large-area Mura defect datasets. As can be seen from the results:

Figure 4 Visual comparison of Mura defect detection results under the black screen scenario.

Figure 5 Visual comparison of Mura defect detection results under the white screen scenario.

Qualitative results. In the Mura defect detection task, accurately segmenting defective regions is critical for improving quality control in display panel manufacturing. Figures 4 and 5 illustrate the detection results under two different background conditions (black screen and white screen), comparing various methods. As can be seen from the results:

Figure 6 Visualization of segmentation, attention, and edge prediction across Mura defect types, demonstrating the effectiveness of the proposed multi-stage fusion and edge-aware architecture.

Overall, unlike other methods that exhibit inconsistent performance across different background scenarios, SIFNet demonstrates consistently strong results, accurately capturing Mura defect features with strong robustness.

4.3 Ablation Studies

To analyze the contribution of each lightweight module in our approach to Mura defect detection, we evaluate four variants of the model: No.1. Baseline (consists of the MobileNetV2 backbone and our lightweight decoder); No.2. Baseline + NEEM; No.3. Baseline + GSIB; No.4. Baseline + NEEM + GSIB (our final model). The quantitative results and computational complexity on the Uneven_Mix dataset are summarized in Table 4. As shown, each module contributes positively to the overall performance of the model in detecting Mura defects. The baseline model (No.1), which consists of a MobileNetV2 backbone and a lightweight decoder, achieves a reasonable balance between speed and accuracy. However, introducing NEEM (No.2) results in improved pixel-level performance, especially in the $F_m^{mean}$ score (0.887) and object-level precision (0.818), confirming the effectiveness of NEEM in refining local defect boundaries. However, the GSIB-enhanced model (No.3) offers even greater improvements in both object-level (precision: 0.831, recall: 0.804) and pixel-level ($F_m^{mean}$: 0.908, recall: 0.611) performance, highlighting its capability to capture large-scale brightness inconsistencies. Finally, the integrated model (No.4), combining both NEEM and GSIB, achieves the best performance across almost all metrics, including the highest object-level precision (0.859) and recall (0.823), highest pixel-level precision (0.824) and $F_m^{mean}$ (0.632), and the lowest mean absolute error (M = 0.042). These results confirm the complementary strengths of the two modules: GSIB enhances global context awareness, while NEEM improves fine-grained boundary localization. Their integration leads to a robust and accurate Mura defect detection framework, demonstrating the effectiveness of our modular design.

4.4 Feature visualization

Figure 6 illustrates the comprehensive visual analysis of the proposed method across various Mura defect scenarios. From top to bottom, the rows correspond to the original input images, ground truth (GT) segmentation masks, predicted segmentation outputs (Prediction), attention activation maps from three decoders (De1_attn, De2_attn, De3_attn), ground truth edge maps (Edge_GT), and the corresponding predicted edges (Edge_Pre). The segmentation results demonstrate high consistency with the ground truth, indicating the model's strong capability in accurately localizing Mura defects across different background conditions and defect morphologies. The attention maps reveal a progressive refinement of spatial focus through the decoder hierarchy, with deeper stages (De2 and De3) exhibiting enhanced sensitivity to defect contours and intensity transitions. Furthermore, the predicted edge maps show a close correspondence to the ground truth boundaries, validating the efficacy of the proposed edge-guided refinement strategy in enhancing boundary localization accuracy. Collectively, these visualizations substantiate the effectiveness of our attention-enhanced multi-stage decoder and edge-aware learning framework in achieving fine-grained, structurally coherent Mura defect segmentation.

Figure 7 Some failure cases.

4.5 Failure Case

Although our model outperforms state-of-the-art (SOTA) methods in both qualitative and quantitative evaluations, several cases where detection results are less satisfactory are illustrated in Figure 7. Specifically, the first row presents a smooth image without visible defects, consistent with the ground truth. However, the prediction map exhibits some uncertain responses, suggesting that the model may be less reliable in defect-free white screen scenarios. This limitation could potentially be addressed by expanding the annotated white screen dataset. In the second row, slight defects are present and correctly labeled in the ground truth. The model detects part of the defective area but does not achieve full coverage. The third row shows clear defects, which the model identifies well, closely matching the ground truth, although minor omissions or over-detections remain. Enhancing image contrast may help highlight Mura regions and further improve detection accuracy. Moreover, Table 5 compares the performance of SIFNet on the original Uneven_Black dataset versus the augmented Black dataset (Black_aug; 1,500 images), reporting improvements across all evaluated metrics: Precision (P), Recall (R), S-measure ($S_m$), mean F-measure ($F_{\beta }^{mean}$), mean E-measure ($E_{\varphi }^{mean}$), max F-measure ($F_{\beta }^{max}$), max E-measure ($E_{\varphi }^{max}$), adaptive F-measure ($F_{\beta }^{adp}$), adaptive E-measure ($E_{\varphi }^{adp}$), and mean absolute error (M). The results demonstrate SIFNet's consistent gains on the augmented dataset while preserving computational efficiency for industrial applications.