How to fuse multi-scale information o f image in deep learning is a key problem to be solved.To solve these problems,this paper proposed a deep learning method base d on multi-scale iterative training for image semantic understanding.The algorithm uses the convolution neu ral network (CNN) to extract dense feature vectors from raw pixel for encoding regions centered on each pixel.The multi-scale iterative training captures different scales of textures,colors,edges and other important informa tion.A new method combined with superpixel segmentation is proposed,to estimate the leading category of superpi xel block and to correct the pixel classification error.It can depict the outline of the target area boundary a nd complete the final semantic understanding.The experiments on Stanford Background Dataset-8verify the effectiveness of the proposed method,and the accuracy rate is 77.4%.