The invention provides a video recommendation method based on a user head portrait and a video recommendation system based on user head portrait. Combined training is performed according to the standardized head portraits of massive users and the watched video data so as to obtain a convolutional neural network model based on image representation. The possibly interested videos of the users can beautomatically learnt according to the head portraits uploaded by the users so that the problem of video recommendation of the cold start users can be effectively solved. Meanwhile, feature vectors obtained by vectorization processing of the semantic vectors outputted by the trained convolutional neural network model and the basic information of the users and the feature information of the presetmultiple candidate recommendation videos are inputted to the recommendation system to be processed. The machine learning result based on the user images, the basic information of the users and the features of the candidate recommendation videos are simultaneously considered so that the accuracy and the fineness of the recommendation videos for the users finally obtained by the recommendation system can be guaranteed.