A Genetic Algorithm-Based Ensemble Convolutional Neural Networks for Defect Recognition with Small-Scale Samples

In modern manufacturing, defect recognition is an important technology, and using recent advances, such as convolutional neural networks (CNNs) to help defect recognition have addressed many attentions. However, CNN requires large-scale samples for training. In industries, large-scale samples are usually unavailable, and this impedes the wide application of CNNs. Ensemble learning might be a feasible manner for the small-scale-sample problem, But the weight for different CNNs needs explicit selection, and this is complex and time-consuming. To overcome this problem, this paper proposes a genetic algorithm (GA)-based ensemble CNNs for small-scale sample defect recognition problem. The proposed method uses an ensemble strategy to combinate several CNN models to solve the small-scale-sample problem in defect recognition, and use GA to optimize the ensemble weights with 5-fold cross-validation. With these improvements, the proposed method can find the optimal ensemble weight automatically, and it avoids the complex and explicit parameter selection. The experimental results with different trainable samples indicate that the proposed method outperforms the other defect recognition methods, which indicates that the proposed method is effective for small-scale sample defect recognition tasks. Furthermore, the discussion results also suggest that the proposed method is robust for noise, and it indicates that the proposed method has good potential in defect recognition tasks.