Fuzzing is widely used for different kinds of software and systems to detect the vulnerabilities. The effectiveness and efficiency of fuzzing is related to the mutation strategy of the seed files and the code coverage of the seed files for the target program. This study proposes a new method based on deep learning for seed generation. The proposed method analyses and learns the correlation between the seed files and their paths in the target program. Finally, the proposed method generates seed files that more likely explore uncovered paths, thus increases the code coverage of the initial seed files for the target program. Aiming at the PDF reader, we carry out the experiment. The results demonstrates that the seed files generated by proposed method have a good passing rate of the PDF reader, in the meantime, significantly improve the code coverage. The experiment also indicates the applicability of proposed method:the seed files which are generated for specific target program (PDF reader) can also obtain higher code coverage when fuzzing some other kinds of PDF readers.