Using visual detection technologies based on deep learning to identify the maturity of papaya fruits on tree in natural environment and monitor the growing periods of papaya is of great significance to the intelligent management of papaya orchard. At present, there are relatively few studies on the identification of papaya maturity. The maturity of papaya is mainly judged manually, which have urgent needs to be replaced by some alternative fast and accurate automatic detection methods. Based on the lightweight YOLO v5-Lite model, a method of papaya maturity detection in natural environment was studied. The detection algorithm was improved based on the YOLO v5 network. In order to alleviate frequent slicing operations, a faster convolution operation was used to replace the Focus layer of the original network, which reduced the amount of computation and released the memory usage and accelerated the inference speed. To reduce the amount of calculation, the ShuffleNetv2 was used in the model to change the 1×1 group convolution in the middle to ordinary convolution by reducing the use of group convolution. At the same time, the ordinary convolution on the branch was changed to a depth-wise separable convolution, which greatly reduced the amount of calculation and improved the calculation efficiency. The number of C3 Layers especially the ones in deep neural blocks was reduced, so as to reduce the cache space occupation and speed up the operations. The channel number in FPN and PAN was set identical to speed the memory accessment. Totally 1386 papaya images were selected to create a dataset in PASCAL VOC format. Under the Ubuntu 16.04 environment, the training parameters of network were set as the epoch number of 300, the batchsize of 128, the total number of iterations of 300, and the initial learning rate of 0.001. During training, the loss value of the model tended to stabilize at the 200th iteration, indicating that the network was converged and the training performance was good. The evaluation indicators for papaya maturity identification of the experiments were the accuracy rate, recall rate, overall average accuracy, detection speed and model size. The experimental results showed that the mAP of the papaya maturity detection model was 92.4%, which outperformed the mainstream lightweight object detection algorithms namely the YOLO v5s and the YOLO v4-tiny and the classic two-stage algorithm Faster R-CNN by 1.1 percentage points, 5.1 percentage points and 4.7 percentage points, on mAP, respectively. In addition, under the condition of relatively accurate detection, the detection time was up to 7ms, and the model size was only 11.3MB. At the same time, the model can accurately identify the fruits under different shooting distances, occlusion conditions and lighting conditions, showing the performance of fast and effective identification and good robustness under complex backgrounds. The proposed method provided technical support for yield estimation of papaya orchards and the positioning detection of picking robots, which can also provide reference for researches on the maturity detection of other fruits.