Personalized picture retrieval method in federal learning scene

The invention discloses a personalized picture retrieval method in a federated learning scene. Comprising the following steps: 1) loading a public picture-text pair data set to generate a model training data set; 2) based on the text information in the data set in the step 1), obtaining a position mapping relationship between words and Token through a token word segmentation device; 3) performing preprocessing based on the picture information in the data set in the step 1) to obtain a picture tensor; and 4) constructing a unified model based on the pre-training model CLIP. 5) training a unified model by using the text and image expressions obtained in the steps 2) and 3); and 6) performing image retrieval by using the unified model trained in the step 5) to obtain a picture conforming to the query text. According to the method, the private retrieval database is designed for the federated learning client, and information specific to the client is fused for the picture and the text, so that the problems of redu