Fig1.
The network framework of the proposed model.
represents the foggy image,
is the dehazed image of
,
is the clear image corresponding to
, and
is the final object detection result.
Fig 2.
The patchwise contrastive learning method based on PatchNCE loss.
Fig 3.
The improved network diagram of YOLOv9s.
Fig 4.
The efficient multi-scale attention.
Fig 5.
IoU measures the overlap between predicted bounding boxes and ground truth boxes.
It is defined as the intersection area divided by the union area of the two boxes. WIoU implements a dynamic, non-monotonic focusing mechanism by IoU.
Table 1.
Results on the hazed COCO2017 dataset.
Table 2.
Results on the hazed RTTS dataset.
Fig 6.
The confusion matrix of our model in the hazed COCO2017 dataset.
Table 3.
Model performance.
Table 4.
Image dehazing.
Table 5.
Influence of PatchLoss.
Table 6.
Comparison among different attention mechanisms.
Table 7.
Influence of EMA attention.
Table 8.
Influence of Wise-IoU.
Fig 7.
Training results under different parameters.