MD Loss: Efficient Training of 3D Seismic Fault Segmentation Network under Sparse Labels by Weakening Anomaly Annotation (2110.05319v6)

Abstract: Data-driven fault detection has been regarded as a 3D image segmentation task. The models trained from synthetic data are difficult to generalize in some surveys. Recently, training 3D fault segmentation using sparse manual 2D slices is thought to yield promising results, but manual labeling has many false negative labels (abnormal annotations), which is detrimental to training and consequently to detection performance. Motivated to train 3D fault segmentation networks under sparse 2D labels while suppressing false negative labels, we analyze the training process gradient and propose the Mask Dice (MD) loss. Moreover, the fault is an edge feature, and current encoder-decoder architectures widely used for fault detection (e.g., U-shape network) are not conducive to edge representation. Consequently, Fault-Net is proposed, which is designed for the characteristics of faults, employs high-resolution propagation features, and embeds MultiScale Compression Fusion block to fuse multi-scale information, which allows the edge information to be fully preserved during propagation and fusion, thus enabling advanced performance via few computational resources. Experimental demonstrates that MD loss supports the inclusion of human experience in training and suppresses false negative labels therein, enabling baseline models to improve performance and generalize to more surveys. Fault-Net is capable to provide a more stable and reliable interpretation of faults, it uses extremely low computational resources and inference is significantly faster than other models. Our method indicates optimal performance in comparison with several mainstream methods.

• The paper introduces MD Loss and Fault-Net to efficiently train 3D seismic fault segmentation networks using sparse labels.
• MD Loss is designed to mitigate the negative impact of false negative labels prevalent in sparse manual annotations, improving model performance.
• Fault-Net is an efficient network architecture that utilizes edge features for better detection quality with significantly fewer parameters and operations.

Overview of "MD Loss: Efficient Training of 3D Seismic Fault Segmentation Network under Sparse Labels by Weakening Anomaly Annotation"

The paper presented by Dou et al. introduces MD Loss (Mask Dice Loss), a novel approach to improve the training of 3D seismic fault segmentation networks when limited to sparse labeling conditions. This method is geared towards suppressing the negative impact of false negative labels prevalent in manual annotation, which is often rife with faults characterized by edge features in seismic data. Additionally, the paper introduces Fault-Net, an innovative network architecture designed to enhance the detection quality and computational efficiency for seismic fault segmentation.

Key Contributions

1. Mask Dice (MD) Loss:
   • MD Loss offers a mechanism to incorporate limited 2D manual annotations into the training of 3D models effectively. It mitigates the influence of false negatives, thus enhancing model performance across various geological surveys.
   • The mathematical formulation and theoretical grounding show that MD Loss significantly reduces the gradient errors induced by false negative labels, which is a common issue in manually labeled data.
   • By employing a region-based loss function, it allows for less sensitivity to incorrect labeling, unlike conventional distribution-based losses such as Binary Cross-Entropy (BCE).
2. Fault-Net Architecture:
   • Fault-Net is designed to capitalize on the edge characteristics of seismic faults, utilizing high-resolution feature representations which favor scalability and efficiency.
   • Integration of Multi-Scale Compression Fusion (MCF) blocks ensures preservation of edge information during feature fusion, leading to improved detection reliability without an extensive demand for computational resources.
   • This lightweight network paradigm demonstrates a substantial reduction in parameters (about 0.42M versus 6.55M in some larger networks) and significantly fewer floating-point operations (FLOPs), making it exceptionally efficient for practical applications.

Experimental Validation

• The experimental setup involved datasets comprising both synthetic and field data from various surveys, allowing for extensive testing under real-world conditions.
• Notably, using MD Loss has yielded consistent improvements, showing superior generalization across surveys, notably outperforming BCE-based approaches by mitigating false negative impacts.
• Quantitatively, MD Loss variants showed nuanced improvements, with parameter $\gamma$ adjustments further optimizing the suppression of false negatives, demonstrating optimal performance at $\gamma=0.7$.
• Qualitative assessments across multiple seismic survey data also indicated marked superiority in fault plane continuity and less noise interference when using Fault-Net.

Broader Implications

The introduction of MD Loss and Fault-Net has significant implications for seismic fault detection in geophysical exploration. By allowing robust training under sparse labels, this method reduces the burden of exhaustive manual annotations while maintaining model performance and reliability. The computational efficiency of Fault-Net further promises pragmatic deployment in industry settings without prohibitive resource costs, facilitating more accessible and rapid fault detection.

Future Directions

This research sets the stage for further exploration into hybrid training paradigms that seamlessly integrate human expertise with data-driven approaches. As AI continues to evolve, augmenting networks with adaptive learning mechanisms that more dynamically suppress noise and false detections could further evolve seismic interpretation methodologies. Additionally, the enhancement of network architectures to balance parameter efficiency and detection accuracy will be vital in adapting these models to more diverse and complex geological settings.