Detecting bagged citrus using a Semi-Supervised SPM-YOLOv5

Abstract: Bagged citrus has triggered a dramatic decrease in the texture details, due to the shape evolution from round to stripping during processing. A great challenge has also been posed in citrus object detection, depending mainly on the number of labeled samples. In this study, an accurate and rapid detection method was proposed for the bagged citrus based on the Teacher-student model with a Strip Pooling Module (SPM)-YOLOv5 algorithm. The images of bagged citrus were collected in the Paidengte Agricultural Science and Technology Demonstration Park, Bishan District, Chongqing of China. The data set of bagged citrus was generated by the image cleaning, enhancement, and labelling tags. Firstly, the stripe attention module was added to the backbone network of YOLOv5. Much attention of the model was drawn to the striped bagged citrus and branches, in order to reduce the average pooling focus on a large number of unrelated areas. Besides, the SPM was integrated into the backbone network of YOLOv5. Among them, the horizontal and vertical pooling were focused on the encoding remote context along the horizontal or vertical spatial dimensions. The SPM was also used to solve the overlapping of each spatial position in the feature map. Specifically, the global horizontal and vertical information was encoded to balance the own weight for the feature modification, in order to effectively expand the receptive field of the backbone network. As such, the SPM was different from the global pooling that only focused on one area. The striped pooling was utilized to easily realize the characteristics of discrete distribution in the horizontal and vertical pooling for a long time. The stripe kernel was used for the feature extraction in the horizontal and vertical directions, in order to capture more local details in the stripe pooling. In doing so, the strip pooling was different from the traditional space pooling that depended on the square core. At the same time, the Teacher-student model was semi-supervised to calculate the consistency loss for the unlabeled samples. Two stages were mainly composed of the model. The first stage was Burn-In. The effective pseudo tags were generated for the teacher model to be well initialized. Therefore, the teacher model was then initialized with the labeled samples. The second stage was mutual learning between the teachers and students. The model was trained using the labeled and unlabeled samples. The robustness of the model was enhanced to reduce the consistency loss in the training process. The target detection was performed on the unlabeled samples, in order to improve the performance of the model and reduce the dependence on labeled samples. The experimental results demonstrated that the average precision of SPM-YOLOv5 for the bagged citrus and branch detection was 77.4% and 53.5%, respectively, which was 7.5% and 7.6% higher than that of YOLOv5. The precision and recall rate of bagged citrus detection reached 94% and 76.2%, respectively. More importantly, the precision of SPM-YOLOv5 based on the Teacher-student model reached 82.6% under the condition of occlusion and overlapping. Meanwhile, the best detection was achieved in 1500 unlabeled and 500 labeled samples. Therefore, the SPM-YOLOv5 based on the Teacher-student model can be expected to detect bagged citrus with higher accuracy and faster speed than before.