Fig. 1. Network structure diagram
Fig. 2. Receptive area mapping
Fig. 3. Pixel-level denoising diagram
Fig. 4. Pixel-level denoising effect
Fig. 5. Semantic enhanced structure
Fig. 6. Asymmetric convolutional layer structure
Fig. 7. Structure of Transformer Encoder
Fig. 8. Different backbones network detection results
Fig. 9. Different modules detection effect
Fig. 10. P-R diagram of different comparison methods
Fig. 11. Different methods detection effect
Fig. 12. Different scenes detection effect diagram
Datasets | Sensors | Resolution/m | Polarization | Scenes |
---|
SSDD | RadarSat-2 TerraSAR-X Sentinel-1 | 1~5 | HH,HV,VV,VH | Offshore,Inshore, Coherent speckle noise | SAR-Ship-Dataset [17] | Sentinel-1 GF-3 | 3,5,8,10,25 | HH,HV,VV,VH | HRSID [18] | Sentinel-1B TerraSAR-X TanDEM | 0.5,1,3 | HH,HV,VV,VH |
|
Table 1. Data set parameters
Backbone | Recall/% | Precision/% | Average precision/% |
---|
Resnet50 | 97.07 | 95.13 | 96.53 | Resnet101 | 97.58 | 96.11 | 96.73 |
|
Table 2. Detection performance of different backbone networks
Pixel-level noise reduction | Semantic enhancement | Recall/% | Precision/% | Average precision/% |
---|
× | × | 89.94 | 94.26 | 89.23 | √ | × | 96.94 | 96.33 | 96.44 | × | √ | 96.93 | 95.84 | 96.40 | √ | √ | 97.58 | 96.11 | 96.73 |
|
Table 3. Detection performance of different models
Pixel-level noise reduction | Semantic enhancement | Params/M | Training time /FPS | Testing time /FPS |
---|
× | × | 152.76 | 6.89 | 22.64 | √ | × | 160.84 | 6.49 | 20.50 | × | √ | 151.40 | 6.10 | 20.92 | √ | √ | 159.49 | 5.99 | 21.09 |
|
Table 4. Model performance analysis
Methods | Backbone | Recall/% | Precision/% | Average precision/% |
---|
Faster R-CNN | Resnet101 | 88.54 | 88.09 | 87.26 | Retinanet | Resnet101 | 89.17 | 87.94 | 86.91 | FPN | Resnet101 | 92.22 | 86.41 | 90.69 | YOLOv4 | CSPDarknet53 | - | - | 95.60 | WANG[24] | ShuffleNetV2 | - | - | 94.70 | R2FA-Det[25] | Resnet50 | - | - | 94.72 | FBR-Net [26] | - | 92.79 | 94.01 | 94.10 | Ours | Resnet101 | 97.58 | 96.11 | 96.73 |
|
Table 5. Detection performance of different methods on SSDD
Methods | Backbone | Recall/% | Precision/% | Average precision /% |
---|
Faster R-CNN | Resnet101 | 80.24 | 81.39 | 76.35 | Retinanet | Resnet101 | 72.17 | 90.62 | 71.03 | FPN | Resnet101 | 83.40 | 90.80 | 81.38 | Ours | Resnet101 | 91.86 | 81.45 | 88.86 |
|
Table 6. Detection performance of different methods on SAR-Ship-Dataset
Scenes | Methods | Recall/% | Precision/% | Average precision/% |
---|
Inshore | Baseline | 79.01 | 87.07 | 77.79 | Our | 91.98 | 89.76 | 90.00 | Offshore | Baseline | 92.78 | 96.01 | 92.24 | Our | 99.04 | 97.78 | 98.53 | Small-scale target | Baseline | 90.79 | 90.78 | 90.66 | Our | 97.60 | 96.78 | 96.85 | Large-scale target | Baseline | 89.84 | 94.65 | 89.13 | Our | 97.37 | 92.50 | 94.64 |
|
Table 7. Detection performance of different scenes