In lensfree on-chip microscopy, the iterative phase retrieval with defocused images easily enables a high-resolution and whole field reconstruction. However, on the reconstruction of the dense sample, conventional methods suffer from the stagnation problem and noise affection under two intensity measurements, which gives rise to a remarkable loss of the image contrast and resolution. Here we propose a novel dual-plane phase retrieval algorithm to perform a stable and versatile lensless reconstruction. A weighted feedback constraint was utilized to speed up the convergence. Then, a gradient descent minimization based on total variation metric was proposed to suppress the noise affection. With these two object constraints, a smoothed but resolution-preserving result can be achieved. Numerical simulations of Gaussian and Poisson noise were given to prove the noise-robustness of our method. The experiments of USAF resolution target, H&E stained pathological slide, and label-free microglia cell demonstrated the superior performance of our approach compared to other state-of-the-art methods.