The image segmentation based on Markov Random Field (MRF) tries to find the maximum a posterior (MAP) global optimal solution, which describes image data relations by local correlations. Comparing with the Simulated Annealing (SA) that is used in the canonical MRF, Genetic Algorithm (GA) has been applied into the optimization computation. Currently the weights of energy function and conditional probability are adjusted by generations' number, which converged so quick that the roles of conditional probability are nearly negligible. On the other hand, many scholars defined the individuals of GA as a set of pixels with the gray-level value, which cause the algorithm sensitivities to the noise and the confusions of gray-level information. This paper presented an improved GA-MRF with dynamic weights. The improved GA-MRF defined the labels coding in a neighborhood as an individual instead of the gray-level values coding in a neighborhood. Furthermore the mechanism of dynamic weights is introduced into the process of optimizing, which balanced the roles between MRF potential energy and conditional probability. The followed Synthetic Aperture Radar (SAR) images segmentations experiments proved that the improved GA-MRF with dynamic weight could reach a satisfied result and avoid trapping into the over-optimizing by MRF models.