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

TL;DR

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 multiscales information, which allows the edge information to be fully preserved during propagation and fusion, thus enabling advanced performance via few computational resources.

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.

Figures (13)

• Figure 1: Most of the faults in the seismic image are perpendicular to the inline direction of the slice, so when labelling faults, the slice is generally made along the inline direction, but some data have crisscross faults, which may result in the slice being parallel to the fault (green box), and slices that are parallel to the faults are difficult to observe with the human eye (yellow box), which is one of the reasons for FNL. There are also certain faults that are very difficult to annotation manually even if they are not parallel to the slice.
• Figure 2: The values $\mu$ at the same position in different channels on the feature map share the parameter $\{w_1,w_1,...w_c\}$ of conv$_p$. The key to using sparse labels is to find the valid gradient about $\{w_1,w_1,...w_c\}$ using the loss function, and we obtain the MD loss by incorporating the mask mechanism in the dice loss and deriving its differentiable form.
• Figure 3: The model has the ability to segment faults, which can be regarded as the later stage of model training. (d) shows that the $\lambda$-BCE is very sensitive to FNL labels and (e) shows that the response of MD loss to FNL is very low in the late training period.
• Figure 4: Fault-Net Structure. Although we also use low-resolution features, not by concatenation, but by adding two low-resolution branches while preserving the high-resolution features. Thus Fault-Net always keeps the high-resolution propagation features after two downsampling, which allows the edges to be fully preserved. Also, when it is necessary to fuse features of different resolutions (multi-scales), we embed the proposed MCF block so that the edge features are not distorted during fusion.
• Figure 5: Structure of MCF block. Different from the convolutional fusion with fixed weights, we decouple the convolutional fusion process into two branches of feature selection and channel fusion, separate the weighted part from the convolution, and generate weights adaptively according to the input to prevent the loss of edge information caused by the fusion.