Picture selection method applied to model retraining

The present invention discloses a picture selection method applied to a model retraining. The model is suitable for judging the component types of a plurality of component pictures, and each component type corresponds to at least one installation site. The picture selection method at least includes the following steps: establishing a picture labeling library; wherein the picture labeling library includes a plurality of component pictures, the component pictures are labeled with a corresponding component type and an installation site respectively; creating an error record; wherein the error record includes at least one wrong component type previously judged wrong by the model and at least one wrong installation site corresponding to the wrong component type; and executing a picture selection assignment to create a data set.