中圖分類號:S126 文獻標識碼:A 文章編號:2095-5553(2025)07-0104-07
Abstract:Toaddress the challenge posedbyvariations inreddate defectsacrossdiffrentvarietiesandlighting conditionsinautomatedsorting tasks,thisstudyproposedanovelmeta-learning-basedalgorithm fordomainadaptivedefect detection.First,across-domaindataset wasconstructedbycolecting images ofreddate defects frommultiplevarietiesand environmental conditions.To mitigatesampleimbalance,aditional defectsampleswere generatedusing the StyleGAN3 network,and data augmentation techniques were applied toenhance the diversity of testdataset.Next,abi-level optimizationmeta-learning framework was introduced fordomain-adaptivereddate defect detection.Aconvolutional neuralnetwork wasemployedasthebaselearner,whileadual-layeroptimizationstrategy wasusedtoconstruct themeta-learner.AnL2regularizationtermwasincorporatedintothelossfunctiontoreduce overfiting.Averageacuracywasused as theevaluationmetric.Ablation experiments wereconductedonboth the base learnerandthe meta-learner,and the proposed methodwas comparedagainst various deep learning and metalearningalgorithms to validate itsperformance.Experimental results demonstrated that theproposedmethod achieves average accuracies of 78.6% on the original target domain dataset and 86.5% on the augmented datasets, outperforming the MAML algorithm by 6.4% and 7.6% ,respectively. These findings confirm the method's effectiveness in adapting to cross-domain red date defect detection under diverse conditions.
Keywords: jujubes defect detection;domain adaptation;meta learning;bi-level optimization;L2 regularization
0 引言
紅棗作為中華民族的代表性果品之一,口感甘甜,含有豐富的營養物質,被廣泛應用于中醫藥和食品制作中。近年來,隨著人們對健康食品關注度的持續攀升,紅棗市場消費量不斷增大,其在食品加工、保健品生產等領域的需求日益旺盛。在紅棗加工過程中,傳統人工分揀方式因效率低下、勞動力成本高且難以保證分揀標準一致性,已無法滿足市場對紅棗供應量和質量的要求。隨著智慧農業的快速發展,果實分揀已逐漸進入智能化應用階段。果實缺陷檢測作為自動化分揀任務中的重要環節,旨在通過圖像處理技術確定果實的缺陷類別,提高分揀的準確性[1]
目前紅棗缺陷檢測方法主要包括基于機器視覺的檢測方法和基于深度學習的檢測方法。機器視覺檢測方法通過提取紅棗圖像特定特征結合傳統分類方法對缺陷紅棗進行識別2,采用人工設計特征提取器的方法往往依賴專家經驗,機器視覺方法對復雜多樣的環境抗干擾性差,例如紅棗的顏色、紋理和形狀可能會受到光照強度和角度的影響,從而導致識別準確率不高。相較于傳統特征提取方法,深度學習算法包含的卷積神經網絡(CNNs)可以進行層次化的特征提取,而且能夠直接從原始數據中學習到有用的特征表示,逐漸成為研究熱點,基于CNNs的深度學習方法在紅棗缺陷檢測任務中已取得了一定的研究進展[3-5]。深度學習模型的訓練通常需要大量的高質量標注數據,在訓練數據和測試數據具有相似分布特征的情況下,其性能表現出色。……