Context-CrackNet: A Context-Aware Framework for Precise Segmentation of Tiny Cracks in Pavement images

TL;DR

Context-CrackNet tackles the precise segmentation of tiny pavement cracks by integrating a Context-Aware Global Module (CAGM) with a Region-Focused Enhancement Module (RFEM) in an encoder–decoder architecture. The CAGM uses linear self-attention to capture long-range global context efficiently, while the RFEM refines skip connections to emphasize fine-grained crack details. Across ten public crack datasets, the framework achieves state-of-the-art performance in mIoU and Dice, with ablation studies confirming the complementary benefits of RFEM and CAGM. The model balances segmentation accuracy with computational efficiency, enabling potential real-time deployment in large-scale pavement monitoring and preventive maintenance systems.

Abstract

The accurate detection and segmentation of pavement distresses, particularly tiny and small cracks, are critical for early intervention and preventive maintenance in transportation infrastructure. Traditional manual inspection methods are labor-intensive and inconsistent, while existing deep learning models struggle with fine-grained segmentation and computational efficiency. To address these challenges, this study proposes Context-CrackNet, a novel encoder-decoder architecture featuring the Region-Focused Enhancement Module (RFEM) and Context-Aware Global Module (CAGM). These innovations enhance the model's ability to capture fine-grained local details and global contextual dependencies, respectively. Context-CrackNet was rigorously evaluated on ten publicly available crack segmentation datasets, covering diverse pavement distress scenarios. The model consistently outperformed 9 state-of-the-art segmentation frameworks, achieving superior performance metrics such as mIoU and Dice score, while maintaining competitive inference efficiency. Ablation studies confirmed the complementary roles of RFEM and CAGM, with notable improvements in mIoU and Dice score when both modules were integrated. Additionally, the model's balance of precision and computational efficiency highlights its potential for real-time deployment in large-scale pavement monitoring systems.

Figures (8)

• Figure 1: Overall Architecture of Context-CrackNet: The proposed framework adopts an encoder-decoder structure with two novel components: the Context-Aware Global Module (CAGM) and the Region-Focused Enhancement Module (RFEM). The ResNet-based encoder extracts hierarchical features $\{F_0, F_1, F_2, F_3, F_4\}$, where $F_3$ is processed by the CAGM to model global contextual relationships and generate the contextualized feature map. The decoder pathway integrates RFEMs at each stage, which refine the skip connections between encoder features and upsampled decoder outputs. This refinement enables effective feature modulation for precise segmentation. Finally, the decoder outputs the predicted segmentation mask $\hat{S}$, capturing fine-grained pavement distress details.
• Figure 2: Context-Aware Global Module (CAGM): The module processes the input feature map $F_3 \in \mathbb{R}^{B \times C \times H \times W}$, reshaping it into a sequence $X \in \mathbb{R}^{B \times N \times C}$, where $N = H \times W$. Using a Linear Self-Attention Mechanism, query ($Q$), key ($K$), and value ($V$) projections generate Global Context Sequences ($Z \in \mathbb{R}^{B \times N \times d_k}$). These sequences are reconstructed into the Contextualized Feature Map ($F_{\text{CAGM}} \in \mathbb{R}^{B \times C \times H \times W}$), embedding global dependencies efficiently.
• Figure 3: Region-Focused Enhancement Module (RFEM): The module refines encoder features $F_{e,l} \in \mathbb{R}^{B \times C_e \times H_e \times W_e}$ and decoder features $F_{d,l+1} \in \mathbb{R}^{B \times C_d \times H_d \times W_d}$ by transforming them into intermediate features $X$ and $G$, respectively, via $1 \times 1$ convolutions. These are combined to compute an Attention Map through element-wise addition, a ReLU activation, and a sigmoid activation. The attention-modulated encoder features are concatenated with the upsampled decoder features and passed through a Conv Block ($\text{Conv} (3 \times 3) \to \text{BatchNorm} \to \text{ReLU}$ repeated twice), producing the refined output feature map $F_{\text{RFEM},l} \in \mathbb{R}^{B \times C_{\text{RFEM}} \times H_{\text{output}} \times W_{\text{output}}}$. This process emphasizes fine-grained details while maintaining contextual relevance.
• Figure 4: Examples of data augmentation techniques applied to crack images and their corresponding masks in the DeepCrack dataset. Augmentations include vertical flip, horizontal flip, and shift-scale-rotate, showcasing the spatial transformations employed to enhance diversity and robustness in the training dataset. The top row illustrates augmented images, while the bottom row presents their respective masks.
• Figure 5: Visual comparison of crack detection results across various datasets. The first column displays the input images, followed by the ground truth masks in the second column. The third column shows the predictions of the proposed Context-CrackNet, while subsequent columns present predictions from comparison models including MAnet, PSPNet, DeepLabV3+, and FPN. Rows correspond to individual datasets (e.g., CRACK500, DeepCrack, CrackTree200, etc.). The red boxes highlight areas where comparison models fail to detect tiny cracks effectively, demonstrating the superior performance of Context-CrackNet in accurately capturing fine-grained crack details.