Research And Application Of Deep Learning Algorithm In Surface Defect Recognition

The concept of Deep Learning is derived from the artificial neural network, the model’s architecture of which usually has multiple layers and the ability to imitate the mammalian nervous system’s representation rules. With the purpose to build a multi-level feature representation, the Deep Learning models process the input sample data with continuous iteration and abstract, using the combination of lower-level features information to obtain the higher-level features representation. Deep Learning is a popular research direction in recent years, which has been applied in numerous fields, such as speech processing, image processing, natural language representation and so on, also has achieved a lot of achievements. On the basis of reviewing the relevant literatures, this paper carries out a research on the using of Deep Learning in the identification field of surface defects. The specific content of the work is as follows:1. This paper gives a detailed introduction of Deep Learning, including the history, basic ideas and characteristics. It also contains a specific description which is focused on the several types of Deep Learning models in their structure and training algorithm. In order to meet the requirements of surface defect identification, this paper builds two sets of data samples used for the experiments, which are solar panel defects and capsule defects.2. This paper focuses on the two mainstream types of model structures in Deep Learning, CNN and DBNs. Those two models have been achieved in MATLAB and applied to surface defect identification. On one side, CNN utilizes the learning information about characteristics of the input data in classification and identification. On the other side, DBNs utilizes the reconstructed images obtained by the template to detect surface defects.3. Since that those two mainstream algorithms above in Deep Learning all have insufficient performance in defect identification. This paper combines the advantages of them, utilizing the concept of convolution RBM which is proposed by Lee H and et al, referring the model structure of CNN, to designs a model structure of Deep CRBM, which is called DCBN for short. This model processes the input date layer by layer and utilizes the higher-layer characteristics to identify the defects. As a result, the experiments show that this model along with its algorithm can obtain a higher recognition rate while applying to surface defect identification. Furthermore, this model shows some versatility on the case that it can be applied to two types of defect samples in solar panels and the capsule.