Fig 1.
The overall architecture of SCP-DETR.
(Compared with RT-DETR-R18, the SCP-DETR replaces the second Fusion module in the neck CCFF with our proposed CO-Fusion, incorporates the additional S2 feature layer processed by SPDConv, and replaces the downsampling operation with PSConv.).
Fig 2.
The detailed architecture of SPDConv with a scaling factor of 2.
Fig 3.
Structural comparison between the original Fusion module (top) and our proposed CO-Fusion module (bottom).
In this paper, the number of RepBlocks is set to 3.
Fig 4.
The structure diagram of CSPOKM.
Here, n represents the number of OKMs used, and to reduce the number of parameters, n is set to 1 in this paper.
Fig 5.
The overall structure of the Omni-Kernel Module.
FFT and IFFT represent the Fast Fourier Transform and its inverse, respectively.
Fig 6.
The overall architecture of the pinwheel-shaped convolutional module.
Here, k represents the size of both the surrounding padding and the convolutional kernel, which is set to 3 in this paper. s denotes the stride size, with a value of 2 in this work.
Fig 7.
The effective receptive field of PSConv when k is 3.
The shades of red represent the effectiveness of the receptive field.
Fig 8.
The six PCB defect categories in the PKU-Market-PCB dataset.
Since the target defects are too small, the images have been magnified and cropped for better visualization of the defects.
Fig 9.
The label distribution of the PKU-Market-PCB dataset; (a) Statistical chart of the six defect categories.
(b) Distribution of bounding box sizes in the dataset. (c) Distribution of object center points relative to the entire image. (d) Aspect ratio distribution of target objects relative to the whole image.
Table 1.
Training parameters.
Table 2.
Ablation study of each component in SCP-DETR.
Table 3.
Performance comparison between SCP-DETR and various mainstream object detection algorithms.
Fig 10.
Comparison of normalized confusion matrices between RT-DETR and SCP-DETR on the PKU-Market-PCB dataset.
Fig 11.
Comparison of SCP-DETR and other models on the PKU-Market-PCB dataset during training in terms of mAP@0.5 (left) and mAP@0.5:0.95 (right).
Fig 12.
The prediction accuracy bar chart of SCP-DETR and other models on the six PCB defect categories in the PKU-Market-PCB dataset.
Fig 13.
Visualization comparison of feature maps for 12 channels after adding modules during the feature extraction stage in RT-DETR and SCP-DETR.
Fig 14.
Heatmap comparison of detection results for the six PCB defect categories between RT-DETR and SCP-DETR networks.
The dark red areas indicate regions that the models focus on.
Fig 15.
Detection results of RT-DETR and SCP-DETR on six PCB defect categories.