Author Affiliations
1Shanghai Institute of Technical Physics, Chinese Academy of Sciences, Shanghai 200083, China2Key Laboratory of Infrared System Detection and Imaging Technology, Chinese Academy of Sciences, Shanghai 200083, China3University of Chinese Academy of Sciences, Beijing 100049, Chinashow less
Fig. 1. Structure of WDSR
Fig. 2. Process of sampling
Fig. 3. Diagrams of multiscale feature extraction networks. (a) Feature extraction network in Faster RCNN; (b) multiscale feature extraction network
Fig. 4. Comparison of target detection results. (a) Before improvement; (b) after improvement
Fig. 5. Digram of SROD
Fig. 6. Impact of resolution on the accuracy of dataset recognition
Fig. 7. Comparison of results. (a) Direct detection results; (b) training with infrared data; (c) after super-resolution reconstruction
Fig. 8. Detection results of super-resolution reconstructed image. (a) Bicubic; (b) ScSR; (c) SRCNN; (d) WDSR
Fig. 9. SROD ship detection results of the entire image
Fig. 10. Comparison of detection results of different methods. (a) Real target; (b) detection result of saliency segmentation; (c) detection result of SROD; (d) detection result of Faster RCNN
Fig. 11. Ship targets of different sizes. (a) Large; (b) medium; (c) small
Layer name | Output size | Config |
---|
Conv1 | 1/2 | 7×7, 64, stride 2 | Conv2_x | 1/4 | 3×3 max pooling , stride 2 | ×3 | Conv3_x | 1/8 | ×4 | Conv4_x | 1/16 | ×23 | Conv5_x | 1/32 | ×3 | Classifier | 1×1 | Average pooling, 1000-dFC, Softmax |
|
Table 1. Structure of ResNet101
Parameter | Value |
---|
Array size | 640×480 | Wavelength range /mm | 7.5~14.0 | Pixel size /mm | 17 | Focal length /mm | 35 | Angular resolution /mrad | 0.68 | Data format | U16(Unsigned 16 bits) | Number of pictures taken when the imager sweeps a line | 5 |
|
Table 2. UAV data parameters
Algorithm | T' | N | ΔT | ΔN | P /% | R /% |
---|
Salience algorithm | 196 | 289 | 52 | 93 | 67.80 | 79.03 | Faster RCNN | 180 | 214 | 68 | 34 | 84.11 | 72.58 | SROD | 202 | 228 | 46 | 26 | 88.59 | 81.45 |
|
Table 3. Detection results
Algorithm | Parameter | Number of detected targets | Total |
---|
Large target | Medium target | Small target |
---|
SROD | T' | 48 | 58 | 96 | 202 | ΔT | 1 | 3 | 42 | 46 | ΔN | 10 | 8 | 8 | 26 | N | 58 | 66 | 104 | 228 | P /% | 82.75 | 87.87 | 92.30 | 88.59 | R /% | 97.91 | 95.08 | 69.56 | 81.45 | Faster RCNN | T' | 43 | 51 | 86 | 180 | ΔT | 6 | 8 | 54 | 68 | ΔN | 19 | 9 | 6 | 34 | N | 62 | 60 | 92 | 214 | P /% | 69.35 | 85 | 93.47 | 84.11 | R /% | 87.75 | 86.44 | 61.42 | 72.58 |
|
Table 4. Target type statistics