Orange recognition method using improved YOLOv3-LITE lightweight neural network

Abstract: Orange recognition is one of the key technologies for automatical fruits picking, delicacy management and orchard yield forecast etc. In order to realize the rapid and accurate identification of orange fruits in natural environment, this paper presents an orange recognition method using improved YOLOv3-LITE lightweight neural network. In the field of fruit-picking robots, the accuracy of bounding box regression for fruit recognition directly determines the success rate of robotic hand-picking. In this paper, the proposed method, which introduces GIoU bounding box regression loss function (GIoU reflects the relationship between the target frame and forecast box) to replace traditional loss function of MSE (mean square error) part of the bounding box regression, effectively improves the accuracy of bounding boxregression. In order to facilitate the migration of the model to mobile terminals, theYOLOv3-LITE lightweight neural network model is proposed. Mobilenet-v2, a lightweight network proposed based on the mobile terminal, can effectively reduce the complexity of the model, and is thus used in this paper to replace the backbone network Darknet-53 in the original network model. The mixup-training method combines 2 pictures into 1 with a certain weight to achieve the effect of target occlusion, reduce the impact of differences between images, improve the generalization ability of the model, and reduce over fitting. In this paper, the mixup-training is combined with transfer learning for the purpose of pre-training. The deep learning framework built in the proposed model is TensorFlow and DarkNet. In order to verify the superiority and feasibility of the proposed model, its recognition effect is tested on test samples with different obscured degrees, comparing with the Faster-RCNN, SSD models and original YOLOv3 network model. The values of F1 and AP are used to assess the differences among models. Test results show that, the model proposed in this paper can get significantly good recognition results. In the data sets of lightly obscured fruit, the F1 value and AP value of the orange recognition model proposed in this paper can reach 95.27% and 92.75%, and the Average IoU is as high as 90.65%. In the data sets of severely obscured fruit, the F1 value and AP value of the orange recognition model proposed in this paper are 91.43% and 89.10%, respectively, and the Average IoU is 83.73%. In all the test sets, the F1 and AP values are 93.69% and 91.13%, respectively, and Average IoU is 87.32%. Compared with the original YOLOv3, the values of F1and AP increase by 1.77% and 2.43%, respectively, and Average IoU increases by 4.72%. The detection speed of orange target can reach 246 frames per second; the inferred speed of a single 416×416 picture on GPU is 16.9ms. The detection speed on the CPU can reach 22 frames per second, and the inferred speed can reach 80.9ms. The memory occupied by the proposed model is 28 MB. Moreover, the pre-training method, combined with the transfer learning and mixup training, reduces the training time and memory consumption of the proposed model. Therefore, the orange recognition method proposed in this paper has the advantages of low model memory footprint, high recognition accuracy and fast recognition speed, etc. It can provide new solutions for orange picking robot and orchard yield forecast, and new ideas for intelligent orange industry as well.