SHARP-Net: A Refined Pyramid Network for Deficiency Segmentation in Culverts and Sewer Pipes

TL;DR

SHARP-Net introduces a refined pyramid-based semantic segmentation architecture tailored for defects in culverts and sewer pipes, combining a bottom-up Inception-like pathway with depthwise separable convolutions and a top-down refinement stream. A key contribution is the integration of Haar-like features via a feature-injection gate, which substantially boosts edge and texture discrimination and accelerates convergence. Empirical results on the Culvert-Sewer Defects dataset and the DeepGlobe Land Cover dataset show that SHARP-Net outperforms state-of-the-art baselines, achieving strong IoU scores while using only about 1.32 million parameters, and that Haar-like features can further improve IoU by up to roughly 23–35% depending on the model, underscoring versatility and practical impact for infrastructure inspection and edge deployment. The work also demonstrates robust generalization and faster training, highlighting potential applicability across diverse semantic segmentation tasks and resource-constrained environments.

Abstract

This paper introduces Semantic Haar-Adaptive Refined Pyramid Network (SHARP-Net), a novel architecture for semantic segmentation. SHARP-Net integrates a bottom-up pathway featuring Inception-like blocks with varying filter sizes (3x3$ and 5x5), parallel max-pooling, and additional spatial detection layers. This design captures multi-scale features and fine structural details. Throughout the network, depth-wise separable convolutions are used to reduce complexity. The top-down pathway of SHARP-Net focuses on generating high-resolution features through upsampling and information fusion using $1\times1$ and $3\times3$ depth-wise separable convolutions. We evaluated our model using our developed challenging Culvert-Sewer Defects dataset and the benchmark DeepGlobe Land Cover dataset. Our experimental evaluation demonstrated the base model's (excluding Haar-like features) effectiveness in handling irregular defect shapes, occlusions, and class imbalances. It outperformed state-of-the-art methods, including U-Net, CBAM U-Net, ASCU-Net, FPN, and SegFormer, achieving average improvements of 14.4% and 12.1% on the Culvert-Sewer Defects and DeepGlobe Land Cover datasets, respectively, with IoU scores of 77.2% and 70.6%. Additionally, the training time was reduced. Furthermore, the integration of carefully selected and fine-tuned Haar-like features enhanced the performance of deep learning models by at least 20%. The proposed SHARP-Net, incorporating Haar-like features, achieved an impressive IoU of 94.75%, representing a 22.74% improvement over the base model. These features were also applied to other deep learning models, showing a 35.0% improvement, proving their versatility and effectiveness. SHARP-Net thus provides a powerful and efficient solution for accurate semantic segmentation in challenging real-world scenarios.

Figures (8)

• Figure 1: Architecture of the proposed SHARP-Net. The input image is progressively filtered and down-sampled by a factor of 2 at each layer in the bottom-up pathway on the left. The top-down pathway on the right performs up-sampling operations to reconstruct a colored-masked image. Haar-like features are injected into the second layer of the bottom-up pathway using feature injection gate shown on the bottom left of the figure.
• Figure 2: Applying Haar-like filters for feature extraction: (a) Original image, (b-f) Filter responses from edge and line detection filters, each extracted using the corresponding Haar-like filters shown below them (h-l). (g) An example of the filter response after applying noise reduction using mask region-based method.
• Figure 3: Comparative segmentation results on the culvert-sewer defects dataset are shown, with the first row illustrating joint problem defects and the second row depicting tree root problems: (a) Original images, (b) Ground truth, (c) U-Net, (d) CBAM U-Net, (e) FPN with ResNet, (f) ASCU-Net, (g) SegFormer, (h) The proposed models.
• Figure 4: Comparative segmentation results on the DeepGlobe Land Cover Classification Dataset, with two samples showing different types of land cover: (a) Original images, (b) Ground truth, (c) FPN with ResNet, (d) U-Net, (e) The proposed model.
• Figure 5: Visual comparisons of SHARP-Net results with varying Haar-like features on sewer-culvert defects: (a) Original images, (b) Ground truth, (c) Base model, (d) Two edge features, (e) Three Haar-like features, and (f) Five Haar-like features yielding the highest quality reconstructions.