Applications in industrial and scientific areas need high-quality multispectral images. However, because of various disadvantages, e.g., adverse environment of sensing, limited spectral bands, etc., multispectral images suffer low contrasts, low signal-noise-ratios, etc. As inputs of applications such as tracking, recognition and so on, multispectral images with low quality may cause those applications to fail. In this paper, aiming to meet practical requirements, we propose an algorithm of efficiently improving miultispectral images. The framework of the proposed method is as follows. According to Retinex theory, a multispectral image can be modeled as multiplicative combination of an illumination component and a reflection component. Typically, an illumination component is of low frequency and determines the dynamic range of pixel intensity, while a reflection component is of high frequency and determines the property of an image. Once we can successfully enhance both an illumination component and a reflection component, which is described later, respectively, we achieve the enhancement of a multispectral image by multiplying two enhanced components. First, we estimate the illumination component based on the principal structures extracted from the multispectral image on a multiple scale, i.e. on a low-level scale, middle-level scale and high-level scale, respectively. The mean value of the three principal structures is used as the estimation of an illumination component. Then the global structure contained in the illumination component, can be obtained by analyzing the corresponding information about its histogram, and is involved to enhance the global contrast and the edge details of the principal structure. Second, with the previously computed illumination component, we can easily derive the reflection component from the multispectral image in pixel-wise division operation. There are adequate image details as well as noise in the reflection component. We suppress the noise and keep the image details by using a non local mean filter. And then we enhance the image details by means of local variances. Finally, multiplying the enhanced illumination component by the filtered reflection component, we enhance the multispectral image. In order to verify the efficiency of our algorithm, experiments are conducted over multispectral image sets including X-ray images, ultraviolet images, well illuminated visible light images, poorly illuminated visible light image, and infrared images. The experimental results show that the proposed algorithm can efficiently remove halo artifacts, well suppress noise and obviously improve local details as well as global contrast. Compared with the state-of-the-art algorithms, the proposed method significantly enhances multispectral images both in vision and in objective analysis by means of information entropy and average gradient.