Many real applications require accurate color reproduction, such as textile, printing, art painting archiving and online product exhibition. The most accurate and informative way to represent a natural object’s color is to use its spectral reflectance. However, the very professional measuring instrument of spectrophotometer has the drawbacks of being expensive and having low measuring resolution and slow measuring speed. As digital cameras have become the most widely used devices for color acquisition, developing accurate reflectance estimation methods from RGB responses has received increasing attention. The aim of spectral reflectance estimation is to build estimation function between RGB tristimulus values and the spectral reflectance vector from training samples. Regression methods have been widely used for this problem. Recent studies have shown that natural objects’ reflectance resides on a lower dimensional submanifold which is embedded in the high-dimensional ambient Euclidean space. Due to the limitation of high dimensional and low training sample size, previous global regression approaches could not exploit the local manifold structure well and are prone to be over-fitting. Local linear regression method can improve the problem of overfitting, but the local learning method is susceptible to the influence of outliers, which will lead to under-fitting. Aiming at this problem, this paper proposes a spectral reflectance estimation method based on locally weighted linear regression, which gives different weights to each local training sample within a k-nearest neighbor constraint. The experimental results show that the method based on locally weighted linear regression can make more effective use of local information, alleviate the over-fitting and under-fitting and reconstruct the spectral reflectance more accurately.