Abstract: As a high-value economic crop, strawberry's automatic picking requires target detection and maturity judgment. Traditional strawberry picking analysis methods mainly use simple image processing methods such as color and size analysis, which has high false alarm rate. In this paper, a two-stage detection network YOLO-ResNeXt is proposed. The Strawberry3000 dataset was created according to the Internet images and the actual farmland photos. On this basis, this paper innovatively used the variational auto-encoder (VAE) to search the network structure quickly, which had high efficiency and good effect on the simple structure search. According to the test results, the algorithm can effectively detect strawberry target and analyze strawberry maturity. Compared with the traditional computer vision algorithm, the accuracy and recall rate are greatly improved, which will effectively promote the development of high-value economic crop picking.