To address the problems of inadequate adaptability and poor visual quality in existing infrared and visible image fusion methods under varying luminance conditions, this paper proposes a fusion method based on Retinex theory. First, the dimension of visible light images is enhanced using an encoder, followed by the decomposition of these images into reflectance and illuminance feature maps, which is consistent with Retinex theory. Second, the reflectance feature is combined with the infrared image feature obtained via the encoder,which enhanced using a structure tensor representation. In addition, convolution kernels with varying sizes are employed to extract multiscale features, which enriches the image's hierarchical information. Finally, the decoder reduces the feature map's dimensionality, and a learnable gamma transform layer is introduced to improve the contrast of the fused image. The model's performance is validated using multiple evaluation metrics on the LLVIP public dataset. The experimental results demonstrate that the proposed method enables adaptive fusion of visible and infrared images under different luminance environments, achieving superior fusion results in terms of both visual perception and quantitative assessment.