With the wide application of intelligent devices and the development of artificial intelligence technology， significant value has been demonstrated by nighttime image enhancement in multiple fields such as visual Internet of Things， autonomous driving， medical imaging， and object detection. Nighttime image enhancement is defined as the process of increasing the brightness， enhancing the contrast and restoring the details of images under insufficient lighting conditions.To address the issues of texture detail loss and noise amplification in existing nighttime image enhancement methods， a nighttime image enhancement method based on a multi-scale fusion network （MSFNet） is proposed in this paper.

Firstly， in order to effectively extract the texture detail features of the dark area of the image， a dark area feature extraction module is introduced at the front end of the backbone network. The global features of the image are extracted through convolution to generate channel attention weights， enhancing the channel features related to the detail information. Secondly， multi-branch network modules are designed in the backbone network， and specific sub-networks are designed for different scales， so that rich multi-scale information can be perceived during the enhancement process and the loss of detail reconstruction at a single scale is avoided. Finally， a dynamic weight allocation mechanism is proposed in this paper， through which the multi-scale features of each sub-network are adaptively fused to enhance the feature expression ability and further improve the image enhancement effect.

On the three paired datasets of LOLv1， LOLv2-real， and LOLv2-syn， the PSNR values of MSFNet were higher than those of Retinexformer by 1.37， 0.72， and 0.50， respectively. The SSIM values on the LOLv1 and LOLv2-syn datasets were also higher than those of Retinexformer by 0.012 and 0.007， respectively. On the three unpaired datasets of LIME， MEF， and NPE， the optimal NIQE values were achieved by MSFNet， with values as low as 3.74， 3.91， and 3.36， respectively. The brightness of the enhanced images was significantly improved， and the details were rendered more clearly. Furthermore， on the publicly available nighttime face detection dataset DARKFACE， higher face detection accuracy was achieved for the enhanced images. The AP values under the three IoU thresholds of 0.5， 0.6， and 0.7 were recorded as 0.264， 0.107， and 0.014， respectively. More detail information was extracted from the images enhanced by MSFNet， thereby enhancing the discrimination between the target and the background and demonstrating the effectiveness of this method in downstream tasks.

MSFNet is proposed to address the the issues of texture detail loss and noise amplification commonly encountered during enhancement. First， a dark area feature extraction module is introduced at the front end of the backbone； channel attention weights are generated via convolutions on image features， and channel features related to fine details are emphasized， thereby enabling effective extraction of low-light texture details. Second， multi-branch network modules are designed， in which three sub-networks in parallel extract image features at different scales， allowing the recovery of global structure and local details in nighttime images. Finally， a dynamic weight allocation mechanism is proposed， by which multi-scale features from the sub-networks are adaptively fused； the feature representation capability is strengthened， and image enhancement performance is further improved. MSFNet is shown by the experimental results to outperform existing methods on multiple datasets. In comparative experiments on night face detection， a significant improvement in night target detection accuracy is achieved by MSFNet， thereby verifying its practical value in downstream visual tasks.