Intelligent Recognition of Goji Berry Pests Using CNN With Multi‐Graphic‐Occlusion Data Augmentation and Multiple Attention Fusion Mechanisms

Archives of Insect Biochemistry and Physiology, Journal Year: 2025, Volume and Issue: 118(4)

Published: April 22, 2025

ABSTRACT Goji berry is an important economic crop, yet pest infestations pose a significant threat to its yield and quality. Traditional identification mainly relies on manual inspection by experts with specialized knowledge, which subjective, time‐consuming, labor‐intensive. To address these issues, this experiment proposes improved convolutional neural network (CNN) for accurate of 17 types goji pests. Firstly, the original data set augmented using multi‐graph‐occlusion augmentation method. Subsequently, imported into CNN training. Based ResNet18 model, new CNN, named GojiNet, constructed embedding multi‐attention fusion modules at appropriate locations. Experimental results demonstrate that GojiNet achieves average recognition accuracy 95.35%, representing 2.60% improvement over network. Notably, compared network, training time model increases only slightly, while size reduced, enhanced. The verifies performance through series evaluation indicators. This study confirms tremendous potential application prospects deep learning in identification, providing referential solution intelligent precise identification.

Language: Английский

Estimating Calibrated Risks Using Focal Loss and Gradient-Boosted Trees for Clinical Risk Prediction

Electronics, Journal Year: 2025, Volume and Issue: 14(9), P. 1838 - 1838

Published: April 30, 2025

Probability calibration and decision threshold selection are fundamental aspects of risk prediction classification, respectively. A strictly proper loss function is used in clinical applications to encourage a model predict calibrated class-posterior probabilities or risks. Recent studies have shown that training with focal can improve the discriminatory power gradient-boosted trees (GBDT) for classification tasks an imbalanced skewed class distribution. However, not function. Therefore, output GBDT trained using accurate estimate true probability. This study aims address issue poor context applications. The methodology utilizes closed-form transformation confidence scores relates minimizer true-class posterior Algorithms based on Bayesian hyperparameter optimization provided choose parameter optimizes calibration, as measured by Brier score metric. We assess how affects optimize balanced accuracy, defined arithmetic mean sensitivity specificity. effectiveness proposed strategy was evaluated lung transplant data extracted from Scientific Registry Transplant Recipients (SRTR) predicting post-transplant cancer. also Behavioral Risk Factor Surveillance System (BRFSS) diabetes status. plots, slope intercept, show approach improves while maintaining same according area under receiver operating characteristics curve (AUROC) H-measure. focal-aware XGBoost achieved AUROC, score, 0.700, 0.128, 0.968 10-year cancer risk, miscalibrated equal AUROC but worse (0.140 1.579). method compared favorably standard cross-entropy (AUROC 0.755 versus 0.736 1-year cancer). Comparable performance observed other models task.

Language: Английский