The work aims to solve such problems as poor robustness and matching accuracy induced by difficultly in adapting to the scaling of matching between affine transformed images in the current image matching algorithm. An image matching algorithm based on improved multi-scale Retinex coupled angle constraint rule was proposed. The bilateral filtering was used to replace the Gauss filter in multi-scale Retinex method. The multi-scale Retinex method was improved to reduce the influence of noise, halo and other factors in the image. Subsequently, the Harris algorithm was introduced to detect the features of the image. The Haar wavelet response vector and principal direction in the circle domain of feature points were obtained to construct the sector with the principal direction as the starting point. The grayscale feature in the sector was extracted to obtain the corresponding feature vectors, thus generating the feature descriptors. The angle constraint rule was established by the angle of feature vectors corresponding to the feature points. Finally, the error matching points were detected by normalized cross correlation function and the matching effects were optimized. The proposed algorithm had better matching accuracy and robust performance than the current image matching algorithm. When the scaling ratio reached 50%, its matching accuracy could still be maintained at around 90.08%. The proposed algorithm still has higher matching accuracy under various geometric attacks, which has good reference value in image processing, information security and other fields.