Although Landweber method is a classical algorithm for image reconstruction in electrical capacitance tomography (ECT) system, the application of Landweber method in multiphase flow detection is limited due to its semiconvergence. To solve this problem, this paper analyzed and proved the semiconvergence of Landweber method mathematically so as to explore its physical properties. Based on this, an improved Landweber method with full and stable convergence was proposed by structuring a compression operator for Landweber. Then the “soft field” characteristic of sensitivity field in ECT sensor and its effect on sensitivity distribution, which is an important basis for image reconstruction, were analyzed. The sensitivity distributions for each typical electrode pair under different flow patterns were given in threedimensional graphics. As ignoring “soft field” characteristic leads to inconsistency of the best iteration times of Landweber for different flow regimes, an adaptive iteration stopping criterion was proposed. The criterion is based on a priori condition obtained by training a number of image samples with least square support vector machine. With this criterion, iteration can be stopped when the image reconstructed is the closest to the actual distribution of multiphase flow. Meanwhile, the effect of “soft field” characteristic on image reconstruction could be compensated. The experimental results prove that the method not only has a stable convergence, but also can improve the image accuracy by 16%~50%. Therefore, the fullconvergence Landweber method can be widely used in none realtime multiphase flow measurement and detection system for its high detection precision.