Fig 1.
Depicts the LDIE-FDNet architecture.
The Backbone establishes a six-tier feature pyramid for multi-scale extraction, while the Neck employs a dynamic hierarchical fusion pyramid for feature refinement. The Head handles localization and classification, with PIoU Loss enhancing geometric characterization of target morphology and spatial configuration.
Table 1.
Information flow and computational behavior.
Fig 2.
Low-light image enhancement network based on multichannel Retinex model.
Fig 3.
Conversion Structure Of GSConv.
Fig 4.
GSConv_C3k2 structure.
Fig 5.
DySample structure.
Fig 6.
Architecture of the SDI module.
For clarity, only the refinement process for the third-level features (l = 3) is depicted. SmoothConv denotes a 3x3 convolution for feature smoothing, and X represents the Hadamard product.
Fig 7.
PIoU conceptual structure.
Fig 8.
Driver fatigue recognition system.
Table 2.
Detailed introduction of YawDD data set.
Table 3.
Details of DMS Datasets.
Table 4.
Ablation experiment of YawDD dataset.
Fig 9.
Model state evaluation on the YawDD dataset.
Fig 10.
Model state evaluation on the DMS dataset.
Table 5.
Ablation experiment of DMS data set.
Table 6.
Comparison of different loss functions.
Fig 11.
Actual improvement effect in complex driving scenarios.
Table 7.
Comparative experiments based on YawDD.
Table 8.
Performance comparison across different hardware platforms on YawDD.
Table 9.
Comparative experiment of DMS.
Table 10.
Performance comparison across different hardware platforms on DMS.