Author Affiliations
1School of Microelectronics, Tianjin University, Tianjin 300072, China2Tianjin Institute of Surveying and Mapping Co., Ltd., Tianjin 300381, Chinashow less
Fig. 1. Architecture of proposed algorithm
Fig. 2. Block diagram of RGB-T image detail information fusion
Fig. 3. Architecture of the proposed FSDNet module
Fig. 4. Salient map comparison between different algorithms. (a) RGB images; (b) T images; (c) EGNet; (d) EGNet+;(e) CPDNet; (f) CPDNet+; (g) PoolNet; (h) PoolNet+; (i) DMRA; (j) A2dele; (k) proposed algorithm; (l) GT
Fig. 5. P-R curves of different algorithms at different datasets. (a) VT821; (b) VT1000
Fig. 6. Qualitative comparison of salient maps for ablation analysis. (a) RGB images; (b) T images; (c) Baseline; (d) +I; (e) +IA; (f) proposed algorithm; (g) -E; (h) GT
Condition | L1 | L2 | L3 | L4 | L5 | L6 | L7 | L8 | L9 |
---|
U | 0.15 | 0.25 | 0.35 | 0.45 | 0.55 | 0.65 | 0.75 | 0.85 | 1 | Q | 0 | 0.15 | 0.25 | 0.35 | 0.45 | 0.55 | 0.65 | 0.75 | 0.85 |
|
Table 1. RGB, T image illumination classification
Parameter | L1 | L2 | L3 | L4 | L5 | L6 | L7 | L8 | L9 |
---|
η | 0.9 | 0.8 | 0.7 | 0.6 | 0.5 | 0.4 | 0.3 | 0.2 | 0.1 |
|
Table 2. RGB, T image light fusion ratio
Parameter | R1 | R2 | R3 | R4 | R5 | R6 |
---|
σ | 2 | 3 | 4 | 5 | 6 | 7 | K | 2 | 4 | 5 | 7 | 8 | 10 |
|
Table 3. Detail information extraction of RGB、T image
L | L1 | L2 | L3 | L4 | L5 | L6 | L7 | L8 | L9 |
---|
D | R6 | R6 | R5 | R4 | R3 | R2 | R2 | R1 | R1 |
|
Table 4. RGB、T Image fusion rules
Algorithm | VT821 | VT1000 |
---|
Fmax | Fave | MAE | Fmax | Fave | MAE |
---|
EGNet | 0.7717 | 0.7459 | 0.0512 | 0.8796 | 0.8637 | 0.0400 | EGNet+ | 0.6994 | 0.6475 | 0.116 | 0.7840 | 0.7453 | 0.0938 | CPDNet | 0.7971 | 0.7882 | 0.0425 | 0.8869 | 0.8621 | 0.0326 | CPDNet+ | 0.7915 | 0.7796 | 0.0471 | 0.8738 | 0.8570 | 0.0333 | PoolNet | 0.8192 | 0.8046 | 0.0485 | 0.8795 | 0.8651 | 0.0411 | PoolNet+ | 0.8177 | 0.8086 | 0.0411 | 0.8759 | 0.8656 | 0.0313 | DMRA | 0.8411 | 0.8270 | 0.0417 | 0.8613 | 0.8444 | 0.0381 | A2dele | 0.7513 | 0.7482 | 0.0621 | 0.8609 | 0.8575 | 0.0401 | CGL | 0.780 | 0.744 | 0.0849 | | 0.727 | | CRA | 0.747 | 0.739 | 0.1083 | | 0.693 | | Proposed algorithm | 0.8433 | 0.8291 | 0.0375 | 0.9095 | 0.8943 | 0.0268 |
|
Table 5. Quantitative comparison of different algorithms
Algorithm | Fmax | Fave | MAE |
---|
EGNet | 0.6602 | 0.5825 | 0.1100 | EGNet+ | 0.8393 | 0.8136 | 0.0616 | EGNet++ | 0.8225 | 0.8036 | 0.0538 | CPDNet | 0.7662 | 0.7092 | 0.1060 | CPDNet+ | 0.8071 | 0.7827 | 0.0721 | CPDNet++ | 0.8164 | 0.8038 | 0.0546 | PoolNet | 0.5803 | 0.5578 | 0.1140 | PoolNet+ | 0.7922 | 0.7683 | 0.0622 | PoolNet++ | 0.7532 | 0.7135 | 0.0940 | Proposed algorithm | 0.8743 | 0.8651 | 0.0386 |
|
Table 6. Test results in VT1000 low light data subset
Algorithm | VT821 | VT1000 |
---|
Fmax | Fave | MAE | Fmax | Fave | MAE |
---|
Baseline | 0.8256 | 0.8003 | 0.0445 | 0.8396 | 0.8315 | 0.0361 | +I | 0.8326 | 0.8125 | 0.0429 | 0.8459 | 0.8361 | 0.0342 | +IA | 0.8493 | 0.8139 | 0.0417 | 0.8956 | 0.8827 | 0.0330 | Proposed algorithm | 0.8433 | 0.8291 | 0.0375 | 0.9095 | 0.8943 | 0.0268 | -E | 0.8366 | 0.8193 | 0.0411 | 0.9051 | 0.8819 | 0.0271 |
|
Table 7. Quantitative comparison results of ablation analysis