Table 2.

Deep-Learning Defect-Detection Methods.

Methods	Strengths	Weaknesses	Applicable
CNN	It has a strong learning ability for high-dimensional input data and can learn abstract, essential and high-order features from a small amount of preprocessed and even the most original data.	The good expression ability and the calculation complex will increase with the increase of network depth.	Unlimited material
Autoencoder neural network	It has a good object information representation ability, can extract the foreground region in the complex background, and has good robustness to the environment noise.	The input and output data dimensions of the autoencoder machine must be consistent.	Unlimited material
Depth residual neural network	The residual network has lower convergence loss and does not overfit, so it has better classification performance.	The network must cooperate with deeper depth to give full play to its structural advantages.	Unlimited material
Full convolution neural network	It can extract the feature of any size image, and obtain the high-level semantic prior knowledge matrix, which has a good effect on semantic level object detection.	The feature matrix transformation combined with the underlying features is needed, and the convergence speed of the model is slow.	Unlimited material
Recurrent neural network	When there are fewer sample data, we can learn the essential features of the data and reduce the loss of data information in the process of pooling.	With the increase of the number of iterations in the network training process, the recurrent neural network model may appear overfitting phenomenon.	Unlimited material