Fig. 1. SSD network structure
Fig. 2. Improvement of SSD network structure
Fig. 3. Different fusion methods
Fig. 4. Transfer learning process
Fig. 5. Cow labeling pictures
Fig. 6. P-R curves of different algorithms
Fig. 7. Detection effect of improved SSD algorithm and SSD algorithm. (a) Improved SSD algorithm; (b) SSD algorithm
Fig. 8. Loss curves of improved SSD algorithms under different training methods on training set
Fig. 9. P-R curves of SSD algorithm with different training methods
Fig. 10. P-R curves of improved SSD algorithm with different training methods
Fig. 11. Detection effect of SSD algorithm with different training methods. (a) Transfer learning trains only the classification layer; (b) new study; (c) transfer learning trains all layers
Fig. 12. Detection effect of improved SSD algorithm under different training methods. (a) Transfer learning trains all layers; (b) new study; (c) transfer learning trains only classification layer
Feature map | M×N | K | Aspect ration | Number of candidate frames |
---|
Conv4_3 | 38×38 | 4 | 1,2 | 5776 | FC7 | 19×19 | 6 | 1,2,3 | 2166 | Conv8_2 | 10×10 | 6 | 1,2,3 | 600 | Conv9_2 | 5×5 | 6 | 1,2,3 | 150 | Conv10_2 | 3×3 | 4 | 1,2 | 36 | Conv11_2 | 1×1 | 4 | 1,2 | 4 |
|
Table 1. Candidate box parameter setting of SSD algorithm
Upper sampling layer | Number of output channels | Kernel_size | Stride | Padding |
---|
concat | add |
---|
Conv11_2 | 256 | 256 | 3 | 3 | Same | Conv10_2 | 256 | 256 | 3 | 1 | Valid | Conv9_2 | 512 | 512 | 3 | 2 | Same | Conv8_2 | 512 | 1024 | 10 | 1 | Valid |
|
Table 2. Parameter setting of upper sampling layer under different fusion modes
Feature map | M×N | K | Aspect_ration | Number of candidate frames | Total number of candidate frames |
---|
FC7 | 19×19 | 8 | 1,2,3,4 | 2888 | | 3948 | Conv8_2 | 10×10 | 8 | 1,2,3,4 | 800 | | Conv9_2 | 5×5 | 8 | 1,2,3,4 | 200 | | Conv10_2 | 3×3 | 6 | 1,2,3 | 54 | | Conv11_2 | 1×1 | 6 | 1,2,3 | 6 | |
|
Table 3. Parameter setting of candidate frame of improved SSD algorithm
Experimental algorithm | Feature fusion | Fusion method | PAP /% | Average detection time /ms |
---|
SSD | | | 88.23 | 46.23 | SSD | √ | Add | 88.11 | 51.52 | SSD | √ | Concat | 90.58 | 54.04 |
|
Table 4. Experimental results on test sets of different fusion methods
Experimental algorithm | NTP/NFP/NFN | P /% | R /% | PAP /% | Average detection time /ms |
---|
SSD | 454/12/60 | 97.42 | 88.32 | 88.23 | 46.23 | Improved SSD | 476/7/38 | 98.55 | 92.60 | 92.55 | 54.16 |
|
Table 5. Experimental results of SSD algorithm and improved SSD algorithm on test set
Experimental algorithm | Training method | NTP/NFP/NFN | P /% | R /% | PAP /% |
---|
SSD | New training | 454/12/60 | 97.42 | 88.32 | 88.23 | SSD | Transfer learning only trains classification layer | 368/3/146 | 99.19 | 71.59 | 71.56 | SSD | Transfer all levels of learning and training | 477/8/37 | 98.35 | 92.80 | 92.69 | Improved SSD | New training | 476/7/38 | 98.55 | 92.60 | 92.55 | Improved SSD | Transfer learning only trains classification layer | 388/11/126 | 97.24 | 75.48 | 75.24 | Improved SSD | Transfer all levels of learning and training | 496/7/18 | 98.60 | 96.49 | 96.40 |
|
Table 6. Experimental results of different training methods on test set
Algorithm | 20% of the training set | 50% of the training set | 80% of the training set | All training set |
---|
SSD | 87.24 | 92.94 | 93.13 | 92.69 | ImprovedSSD | 91.73 | 95.10 | 95.66 | 96.40 |
|
Table 7. Influence of training set size of target domain on recognition accuracy after transfer learning
Algorithm | PAP /% | Average detection time /ms |
---|
SSD | 88.23 | 46.23 | YOLOV2 | 85.26 | 20.49 | YOLOV3 | 90.80 | 53.24 | Method in Ref. [13] | 85.46 | 19.64 | Method in Ref. [14] | 93.21 | 55.01 | Method in this paper | 96.40 | 54.16 |
|
Table 8. Average accuracy of different algorithms