Author Affiliations
1College of Information and Control Engineering, Xi'an University of Architecture and Technology, Xi'an, Shaanxi 710055, China2Guangdong Artificial Intelligence and Digital Economy Laboratory (Guangzhou), Guangzhou, Guangdong 510000, Chinashow less
Fig. 1. Perspective structure of spatial attention
Fig. 2. Multi-scale information aggregation structure
Fig. 3. Asymmetric convolution structure
Fig. 4. Diagram of feature fusion process
Fig. 5. Structural diagram of semantic embedding and feature fusion
Fig. 6. Structural diagram of multi-feature information fusion network
Fig. 7. Flow chart of multi-feature information fusion network algorithm
Fig. 8. Experimental results in ShanghaiTech dataset. (a) Original graphs; (b) true-value graphs; (c) prediction results
Fig. 9. Experimental results in Mall dataset. (a) Original graphs; (b) true-value graphs; (c) prediction results
Method | Part_A | Part_B |
---|
MAE | MSE | MAE | MSE |
---|
Algorithm in Ref.[33] | 181.8 | 277.7 | 32.0 | 49.8 | MCNN[24] | 110.2 | 173.2 | 26.4 | 41.3 | Switch-CNN[34] | 90.4 | 135.0 | 21.6 | 33.4 | MSCNN[35] | 83.8 | 127.4 | 17.7 | 30.2 | CSRNet[7] | 68.2 | 115.0 | 10.6 | 16.0 | SANet[36] | 67.0 | 104.5 | 8.4 | 13.6 | Proposed algorithm | 63.2 | 102.8 | 8.0 | 12.8 |
|
Table 1. Performance comparison among algorithms for ShanghaiTech dataset
Method | MAE | MSE |
---|
Algorithm in Ref. [33] | 3.15 | 15.7 | MCNN[24] | 2.21 | 7.33 | Switch-CNN[34] | 2.01 | 6.25 | MSCNN[35] | 2.12 | 7.04 | DecideNet[37] | 1.52 | 1.90 | Proposed algorithm | 1.43 | 1.72 |
|
Table 2. Performance comparison among algorithms for Mall dataset
Method | S1 | S2 | S3 | S4 | S5 | Average MAE |
---|
MCNN[24] | 3.4 | 20.6 | 12.9 | 13.0 | 8.1 | 11.6 | MSCNN[35] | 7.8 | 15.4 | 14.9 | 11.8 | 5.8 | 11.7 | Switch-CNN[34] | 4.4 | 15.7 | 10.0 | 11.0 | 5.9 | 9.4 | DecideNet[37] | 2.0 | 13.14 | 8.9 | 17.4 | 4.75 | 9.23 | CSRNet[7] | 2.9 | 11.5 | 8.6 | 16.6 | 3.4 | 8.6 | Proposed algorithm | 2.6 | 11.2 | 8.9 | 14.2 | 3.6 | 8.1 |
|
Table 3. Performance comparison among algorithms for Worldexpo’10 dataset
Method | Size /MB | Average running speed of test image /s |
---|
ShanghaiTech | Mall | Worldexpo’10 |
---|
Algorithm in Ref. [33] | 7.1 | 2.36 | 0.32 | 1.22 | MCNN[24] | 19.2 | 2.31 | 0.32 | 1.15 | Switch-CNN[34] | 32.2 | 2.71 | 0.43 | 1.35 | MSCNN[35] | 22.2 | 2.34 | 0.32 | 1.14 | CSRNet[7] | 16.26 | 1.97 | 0.26 | 0.93 | Proposed algorithm (MSIA) | 17.39 | 2.11 | 0.29 | 1.02 | Proposed algorithm ( MSIA+PSA) | 21.4 | 2.32 | 0.31 | 1.11 |
|
Table 4. Comparative analysis of algorithm complexity