Abstract: Conventional dehazing algorithm may cause detail loss, low brightness, and color distortion when it processes a marine image. A dehazing algorithm based on Retinex theory is introduced to solve the problems. The image is transformed into HSV（Hue-Saturation-Value） space before processing. The improved Multi-scale Retinex（MSR） algorithm is used to process V channel and obtain the illumination component. The reflex component is obtained by logarithmic operation. The latter is combined with the normalized H and S channel to obtain the image base layer. The image detail layer is obtained through guided filtering and combined with the base layer for automatic white balance processing to produce final result image. Experiments show that, compared to Dark Channel Prior algorithm and conventional MSR algorithm, the dehazed image produced by the improved MSR is superior in standard deviation, average gradient, peak signal to noise ratio and structural similarity.