Fig. 1. Schematic of depthwise separable convolution

Fig. 2. Schematic of the principle of new partial convolutionh — height of feature map, w — width of feature map, c — number of channels in feature map, c_p — number of imbalanced convolution channels, k — convolution kernel size

Fig. 3. Schematic of novel partial convolution with pointwise convolutionc — number of channels in feature map, c_p — number of imbalanced convolution channels, k — convolution kernel size

Fig. 4. Structural diagram of four-layer backbone feature extraction network structure based on novel partial convolutionConv — convolution, Fasternet — imbalanced convolution core module, c_i — number of input channels, h — height of feature map, w — width of feature map, l_i — number of modules

Fig. 5. Diagram of Fasternet modulePConv — partial convolution, Conv — conventional convolution, BN — batch normalization, ReLu — activation function

Fig. 6. Flowchart of imbalanced depthwise separable convolutional neural network YOLOv8n-WIOU-Fasternet model structureConv — convolution, Fasternet Block — imbalanced convolution core module, SPPF — spatial pyramid pooling fast, C2f — cross stage partial with two convolutions and feature fusion, Concat — concatenation layer

Fig. 7. X-ray raw images of fuel rods

Fig. 8. Photos of typical defects in fuel rods (a) Porosity defect, (b) Tungsten inclusion, (c) Lack of penetration

Fig. 9. Manual tagging of fuel rod defect (a) Projection of a fuel rod assembly within a replenishment block, (b) Single fuel rod projection, (c) Region of interest, (d) Manually labeled defects

Fig. 10. Loss function curves (a) Boundary regression loss, (b) Distribution focal loss

Fig. 11. Curves of false negative ratio-confidence threshold

Fig. 12. Curves of false predicted ratio-confidence threshold

Fig. 13. Curves of false predicted ratio-false negative ratio

Fig. 14. Curves of AP-IoU (color online)