Fig. 1. Network structure of proposed generator

Fig. 2. Residual block in the generator network

Fig. 3. Discriminator network structure

Fig. 4. Different images. (a) River and car of normal-light images; (b) images with illuminance of 0.2; (c) images with illuminance of 0.35; (d) images with illuminance of 0.5; (e) HSV color space images with illuminance of 0.2; (f) H component; (g) S component; (h) V component

Fig. 5. Flow chart of proposed algorithm

Fig. 6. Enhanced results of low-light image Starfish by different algorithms. (a) Low-light image; (b) normal-light image; (c) Ref. [6] algorithm; (d) Ref. [7] algorithm; (e) Ref. [8] algorithm; (f) Ref. [9] algorithm; (g) Ref. [10] algorithm; (h) Ref. [11] algorithm; (i) Ref. [12] algorithm; (j) proposed algorithm

Fig. 7. Enhanced results of low-light image Man by different algorithms. (a) Low-light image; (b) normal-light image; (c) Ref. [6] algorithm; (d) Ref. [7] algorithm; (e) Ref. [8] algorithm; (f) Ref. [9] algorithm; (g) Ref. [10] algorithm; (h) Ref. [11] algorithm; (i) Ref. [12] algorithm; (j) proposed algorithm

Fig. 8. Enhanced results of low-light image Street by different algorithms. (a) Low-light image; (b) normal-light image; (c) Ref. [6] algorithm; (d) Ref. [7] algorithm; (e) Ref. [8] algorithm; (f) Ref. [9] algorithm; (g) Ref. [10] algorithm; (h) Ref. [11] algorithm; (i) Ref. [12] algorithm; (j) proposed algorithm

Fig. 9. Enhancement results of real low-light image Pocky by different algorithms. (a) Real low-light image; (b) Ref. [6] algorithm; (c) Ref. [7] algorithm; (d) Ref. [8] algorithm; (e) Ref. [9] algorithm; (f) Ref. [10] algorithm; (g) Ref. [11] algorithm; (h) Ref. [12] algorithm; (i) proposed algorithm

Fig. 10. Enhancement results of real low-light image Palace by different algorithms. (a) Real low-light image; (b) Ref. [6] algorithm; (c) Ref. [7] algorithm; (d) Ref. [8] algorithm; (e) Ref. [9] algorithm; (f) Ref. [10] algorithm; (g) Ref. [11] algorithm; (h) Ref. [12] algorithm; (i) proposed algorithm

Fig. 11. Comparison results of different algorithms. (a) Average; (b) average gradient

Fig. 12. Subjective comparison of low-light image enhancement results by generator network and generative adversarial network. (a) Low-light image; (b) normal-light image; (c) results of generator network; (d) results of generative adversarial network

Fig. 13. Comparison results of generator network and generative adversarial network of different images. (a) PSNR; (b) SSIM

Table 1. PSNR and SSIM of different algorithms