Fig 1.
Overall Architecture and Inference Pipeline of the MTF-NET Detection Framework.
Fig 2.
Overall architecture diagrams of CAF and CFF.
Fig 3.
Overall architectural diagram of CFF and CAF in MTF-NET.
Fig 4.
Detailed Architecture Diagram of MKDFN.
Fig 5.
Overall structure diagram of HIEP.
Fig 6.
Internal Structure Diagram of HFFU.
Table 1.
Comparison of YOLOv11n and MTF-NET on different classes.
Table 2.
Comparative Analysis of Different Algorithms on the VisDrone2019 Test Set.
Fig 7.
Comparative Analysis of Various Detectors on the VisDrone2019 Dataset.
Table 3.
Comparative experiments evaluating various approaches for small-target enhancement.
Table 4.
Comparative Analysis of the Effectiveness of Different Loss Functions.
Table 5.
Comparative Analysis of Different IoU Functions.
Table 6.
Ablation Study Results on VisDrone2019 Dataset with Various Components and Losses.
Fig 8.
Comparison of detection results between YOLOv11n and MTF-NET under different scenarios.
Rows (a), (b), (c), and (d) correspond to aerial low-light scenes, occlusion scenes, densely populated object scenes, and nighttime scenes, respectively. Columns (e), (f), and (g) represent the original images, detection results from YOLOv11n, and detection results from MTF-NET, respectively.
Fig 9.
Heatmap visualizations of YOLOv11n and MTF-NET under different viewpoints.
Rows (a) and (b) correspond to nighttime scenes and drone-based aerial small-object scenes, respectively. Columns (c), (d), and (e) represent the original images, heatmaps generated by YOLOv11n, and heatmaps generated by MTF-NET, respectively.
Fig 10.
A comparison of detection performance between YOLOv11n and MTF-NET on the UA-DETRAC-G2 dataset (a) and the original HazyDet dataset (b).
Row (a) presents sampled visualized detection outcomes for the UA-DETRAC-G2 dataset, while row (b) illustrates corresponding results for the HazyDet dataset. In both cases, column (c) displays the original images, column (d) presents the detection results from YOLOv11n, and column (e) shows the outcomes achieved by MTF-NET.
Table 7.
Generalization experiments on the UA-DETRAC-G2 dataset.
Table 8.
Generalization experiments conducted on the test set of the original HazyDet dataset.