Abstract Strawberry maturity detection is a key technology for automated strawberry picking and intelligent information monitoring. This paper studies strawberry maturity detection technology, and proposes an improved YOLOv4 convolutional neural network detection method. In view of great amount of network parameters, this paper uses MobileNetv3 backbone feature extraction network and depthwise separable convolution for lightweight improvement of the YOLOv4 network. In order to increase the model training accuracy, this paper uses the Kmeans++ clustering algorithm to calculate the prior bounding box size, and uses transfer learning and staged training methods to improve the training efficiency of the built network model. The experimental results show that the mean average precision (mAP) of the test dataset in this paper is 96.78%; the precision of mature strawberry detection is 98.72%, with recall rate 91.67% and average precision (AP) 99.56%; the precision of immature strawberry detection is 90.76%, with recall rate 83.92% and AP 94.00%. Single image detection time is 56ms, which can meet the demand for real-time high-precision detection of strawberries.