Text extraction from images with complex backgrounds remains a challenging problem. Existing thresholding methods succeed in extracting text from images with simple or slowly varying backgrounds. However, when the backgrounds include sharply varying contours, some background pixels, which have similar intensities to the text, are classified to the text pixels in the binary image. In the literature, seed-fill method is used to remove these background pixels. But, existing seed-fill method cannot remove the background pixels inside the characters. To overcome the disadvantages of the previous methods, we propose a novel text extraction method. This method combines a locally adaptive seed-fill method, a locally adaptive thresholding method and a stroke-model-based method with the following steps: (1) The locally adaptive seed-fill method, the locally adaptive thresholding method and the stroke-model-based method are respectively used to get three binary images; (2) The final binary image is gotten by fusing the three binary images. Experimental results demonstrate the effectiveness of the proposed method in comparison with other related works in the literature.