Abstract:Constructing a large-scale dataset for tea bud object detection is a time-consuming and intricate task. To mitigate the cost of dataset construction, exploring algorithms with a minimal number of annotated samples is particularly necessary. The YOLO singular value decomposition for tea bud detection (YSVD-Tea) algorithm was introduced, which achieved the reconstruction of the YOLOX structure by replacing the basic convolution in the pre-trained model with three consecutive matrix structures. Through dimension transformation and singular value decomposition operations, pre-trained weights were converted into weights corresponding to the reconstructed algorithm structure, thereby separating the weights that require transfer learning from those that needed to be retained. This achieved the goal of preserving the general semantic information of the pre-trained model. Training and validation on three datasets of varying sizes were conducted. On the smallest 1/3 dataset, the YSVD-Tea algorithm showed a 20.3 percentage points improvement in mAP compared with the original YOLOX algorithm. Comparing performance metrics between the test and training sets, the mAP difference for the YSVD-Tea algorithm was only 21.9%, which was significantly lower than YOLOX’s 40.6% and Faster R-CNN’s 55.4%. In training with the largest complete dataset, the YOLOX algorithm achieved precision, recall, F1 score, and mAP of 86.4%, 87.0%, 86.7%, and 88.3%, respectively, surpassing the comparison algorithms. YSVD-Tea algorithm demonstrated superior suitability for the task of tea bud object detection, especially when confronted with a limited number of annotated samples.
