In complex environments with local light sources, the method based on atmospheric scattering model is not accurate enough to estimate atmospheric parameters, and has limitations in dealing with uneven lighting scenes and reflective materials. Therefore, this paper proposes the enhanced haze removal method by fusing improved atmospheric scattering model and Retinex. Firstly, to solve the problem that the traditional atmospheric scattering model ignores the local light source, this paper introduces the space-direction dependent atmospheric light field modeling. We estimate the light source direction through the particle swarm optimization algorithm, construct an anisotropic diffusion equation to solve the light source intensity field, and adjust the scattering coefficient adaptively in combination with the local gradient. This method improves the estimation accuracy of atmospheric parameters in complex scenes. Secondly, aiming at the problems of uneven illumination scenes and the treatment of reflective materials, this paper fuses the Retinex theory to expand the reflectance decomposition model. The observed image is separated into the base reflectance, the illumination modulation factor, and the illumination field. The illumination field reconstruction is realized through an alternating optimization and guided filter. This method effectively deals with the issues of non-uniform illumination and highlight residues. Experimental results show that the proposed method can effectively preserve edge details and improve color authenticity, and it is superior to the contrast algorithm in defogging effect and image quality in complex environment.