A dynamic self-adaptive threshold segmentation algorithm based on human visual attention theory and maximum between-class variance method is proposed.First of all,the pending image is divided into lots of equal-sized square blocks,then the feature value of every block is calculated respectively according to multi-scale visual attention.Secondly,all the image blocks are divided into two types by classification algorithm based on feature value of visual attention.Finally,blocks with different types are binarized by maximum variance algorithm and multi-scale adaptive threshold algorithm respectively.The experimental results show that:compared with the one dimension maximum variance and two dimension maximum variance segmentation methods,this new proposed method can eliminate the disturbance of the uneven lighting and background noises,thus achieve superior segmented results for uneven lighting and poor quality document image binarization.These experimental results testify the validity and practical value of the proposed method.