Abstract
Semantic segmentation of very high-resolution (VHR) remote sensing images is essential for urban planning and environmental monitoring. Traditional deep learning models often struggle with parameter redundancy and limited global context. This paper introduces the Global Evolution Neural Network (GENet), which employs a Knowledge Enhancement strategy to reactivate "dead" convolutional kernels and a Dict-attention module to capture dataset-wide correlations with linear complexity. Evaluated on the ISPRS Vaihingen dataset, GENet achieves state-of-the-art performance, reaching a Mean Intersection over Union (MIoU) of 76.34% and an Overall Accuracy (OA) of 90.72%.
Introduction
The segmentation of VHR images involves assigning class labels to every pixel in high-resolution data. Despite the success of Convolutional Neural Networks (CNNs), they face two major inefficiencies. First, as networks deepen, many kernels become "dead" or redundant, wasting computational resources. Second, standard self-attention mechanisms are computationally expensive and typically only model relationships within a single image, failing to leverage global semantic information across the entire training set.
Methodology
GENet addresses these challenges through two "global evolution" strategies integrated into an encoder-decoder framework.
Knowledge Enhancement
Instead of pruning redundant parameters, GENet reactivates them through mutual learning between two parallel networks ($N_1$ and $N_2$). By measuring information entropy, the model identifies "dead" kernels and transfers knowledge from "valid" kernels in the peer network. This evolutionary process improves feature representation without increasing the computational cost of inference.
Dict-attention
To overcome the $O(N^2)$ complexity of standard self-attention, the authors propose "dict-attention." This module uses shared "dictionary units" to store essential features from the entire dataset. This approach captures cross-image relationships and reduces computational complexity to $O(N)$, making it highly efficient for high-resolution remote sensing data.
System Architecture
The overall architecture of GENet (Figure 1) utilizes a dual ResNet40 backbone to process Digital Orthophoto Maps (DOM) and Digital Surface Models (DSM). It incorporates a Scale-Aware Pyramid Fusion (SAPF) module to effectively capture objects of varying sizes.
Experimental Results
The model was evaluated on the ISPRS Vaihingen dataset, a benchmark for VHR remote sensing containing infrared, red, and green bands along with height data.
Quantitative Performance
GENet outperformed established models such as DeepLabV3+, SCAttNet, and ACNet. As shown in Figure 2, the model achieved the highest MIoU and Overall Accuracy among the compared methods.
Key Insights
- Superior Accuracy: GENet achieved a 76.34% MIoU, with particularly strong results in identifying buildings (88.93% IoU) and trees (70.26% IoU).
- Ablation Study: Removing either Knowledge Enhancement or Dict-attention resulted in a measurable drop in accuracy, proving that both evolutionary steps are necessary for peak performance.
- Visual Quality: Qualitative analysis showed that GENet produces cleaner edges and more complete object segments compared to previous models.
Conclusion
The Global Evolution Neural Network (GENet) successfully mitigates parameter redundancy and enhances global context modeling in remote sensing image segmentation. By reviving "dead" kernels and learning a global feature dictionary, GENet sets a new standard for accuracy and efficiency in the field. Future work will explore applying these strategies to other complex computer vision tasks.
Summary updated