Component platforms are widely used in developing large web and mobile applications to improve development efficiency and test coverage. Searching for the appropriate component in the component platform and introducing it into the project for use makes development efficient and reduces redundant code. Machine learning methods are now widely used in large component platforms to enhance automatic retrieval or identification. However, current machine learning models are less accurate in identifying components or icons with differences due to the difficulty of collecting a large amount of label data in online environments that can help learning. To address this problem, this paper improves the feature mixing method in active learning and embeds it into a classical continuous learning model to solve the problem of accuracy degradation when the model handles discrepant tasks and tasks with limited training data. The model proposed in this paper ensures that the accuracy and performance of the model do not degrade significantly when facing a series of tasks with less data in the dataset after training on the first few tasks. Compared with large neural network recognition tasks, the improved continuous learning model is more lightweight and can learn multiple tasks using a single model. Easy to use when combined with front-end engineering and can be added to projects through dependency libraries. Compared to various classical benchmark models in lifelong machine learning, the accuracy of the proposed model can be improved by about fifteen percent.