When building a classification model, the scenario where the samples of one class are significantly more than those of the other class is called data imbalance. Data imbalance causes the trained classification model to be in favor of the majority class (usually defined as the negative class), which may do harm to the accuracy of the minority class (usually defined as the positive class), and then lead to poor overall performance of the model. A method called MSHR-FCSSVM for solving imbalanced data classification is proposed in this article, which is based on a new hybrid resampling approach (MSHR) and a new fine cost-sensitive support vector machine (CS-SVM) classifier (FCSSVM). The MSHR measures the separability of each negative sample through its Silhouette value calculated by Mahalanobis distance between samples, based on which, the so-called pseudo-negative samples are screened out to generate new positive samples (over-sampling step) through linear interpolation and are deleted finally (under-sampling step). This approach replaces pseudo-negative samples with generated new positive samples one by one to clear up the inter-class overlap on the borderline, without changing the overall scale of the dataset. The FCSSVM is an improved version of the traditional CS-SVM. It considers influences of both the imbalance of sample number and the class distribution on classification simultaneously, and through finely tuning the class cost weights by using the efficient optimization algorithm based on the physical phenomenon of rime-ice (RIME) algorithm with cross-validation accuracy as the fitness function to accurately adjust the classification borderline. To verify the effectiveness of the proposed method, a series of experiments are carried out based on 20 imbalanced datasets including both mildly and extremely imbalanced datasets. The experimental results show that the MSHR-FCSSVM method performs better than the methods for comparison in most cases, and both the MSHR and the FCSSVM played significant roles.

Object detection technologies have played a pivotal role in agricultural informatics at present. Among them, the precision and efficiency of grading systems have been enhanced for tobacco leaves. These technologies can also be integrated to realize the accurate identification and classification of tobacco leaf features. Particularly, the partial occlusion has commonly occurred in the images of multiple tobacco leaves. In this study, an innovative recognition model was proposed using an improved Cascade Mask Region-based Convolutional Neural Network (Cascade Mask R-CNN). Three types of attention mechanisms were also introduced, including channel, nonlocal, and spatial attention. Furthermore, a refined detection framework was then incorporated to employ the SIoU loss function, thereby improving the detection accuracy. The automatic grading of tobacco was effectively realized using visual features. Specifically, Cascade Mask R-CNN was then chosen to treat the complex tasks of image recognition. Then, a series of enhancements were employed to improve the overall performance. Firstly, Channel-Nonlocal-Space (CNS) attention mechanism was introduced into the ResNet101 backbone. The improved model was then focused mainly on the unoccluded regions of the image with distinct and salient features. By doing so, this improved model was used to better discern the nuanced characteristics of tobacco leaves for the classification. Secondly, the conventional SmoothL1Loss function was replaced with the more sophisticated SIoU loss function. The former was typically employed in the Cascade Mask R-CNN for the loss calculation, but without considering the directional difference between the predicted and actual bounding box. As a result, the improved model benefited from the enhanced detection accuracy and expedited convergence, thus facilitating more precise identification of tobacco leaf features. Thirdly, the traditional Non-Max Suppression (NMS) was replaced with the Soft Non-Max Suppression (Soft NMS), in order to avoid the information loss caused by simple and rough deletion of candidate boxes in the standard NMS, thereby preserving more available information for the final analysis. A series of experiments was conducted to verify the effectiveness of the improved model. 1125 images of partially occluded multiple tobacco leaves were acquired from the actual tobacco leaf grading system. An average accuracy of 80.2% (bbox_mAP50) was achieved 7.5 percentage points higher, compared with the Cascade Mask R-CNN model. The better performance was also achieved by 7.1, 10.2, 5.8, 9.2, and 8.4 percentage points, respectively, compared with the multiple mainstream object detection models, namely the YOLOvX, the YOLOv3, the YOLOv5, the Mask R-CNN, and the Cascade R-CNN. Especially, the outstanding performance was also obtained for the lower tobacco leaves difficult to identify. The findings can provide an effective approach for the position recognition of partially occluded multiple tobacco leaves.