In view of the slight deformation in sheet stacks, which leads to the problem of inaccuracy in automatic sheet counting，a new method of automatic plate counting based on skeleton extraction and gradient direction projection statistics was proposed. Firstly, the image was captured with the tablet computer by using the Gabor filter to enhance the direction characteristics of the thin plate edges and extracting the edges by the Sobel operator, and then it was binarized by Mean Shift algorithm, while the interference was removed by being filled with diffuse water, morphology and BlobsFiltering. Finally, a single pixel skeleton was performed on the edge of each plate, and the probability was calculated by the gradient direction projection, thus to realize the accurate counting of the sheets. The experiment was performed on 171 sheets with thickness ranging from 0.8 to 1.2 mm by photographing 10 times. The results showed that the method could reach 100% accuracy and solve the problem in stable automatic counting of non-parallel plates in a certain range.