Fig 1.
Structure of SSD.
Fig 2.
Structure of ISSD.
Fig 3.
Three typical cases of convolution kernel shifting of conventional convolution.
Group A is the conventional distribution, Group B is the distribution after arbitrary migration, Group C is the distribution after scaling transformation and Group D is the distribution after rotational transformation.
Fig 4.
Schematic diagram of depth separable convolution.
Fig 5.
Schematic diagram of the depth separable deformable convolution.
Fig 6.
Schematic diagram of the inception structure.
Group A represents the structure of the original inception and Group B represents the structure of the improved inception.
Fig 7.
Structure of the feature recalibration module.
Table 1.
Number of different samples in the WCD dataset.
Table 2.
Specific index parameters of the workstation.
Fig 8.
Curves of training loss with different learning epochs.
Fig 9.
Comparison of output results at different epochs.
The green boxes locate the missing detection parts of the detection results and the red boxes locate the false detection parts of the detection results.
Table 3.
Network parameter setting.
Table 4.
Validation results of components in the ISSD.
Table 5.
The results of different networks on the WCD dataset.
Table 6.
The computational efficiency and computational complexity of different networks.
Fig 10.
Visualization of detection results of compared networks on the WCD dataset.
Fig 11.
The visualization of detection results of compared networks in real images.