Deep Belief Neworks for Defect Detection
This post explores the use of DBN for image reconstruction and defect detection.
Restricted Boltzmann machines (RBN)
RBM is based on Boltzmann machine. BM is powerful on unsupervised learning, but slow training and impossible to calculate the distribution accurately, and it’s hard to get samples that follow the distribution required by BM. RBM is used to solve those problems.
In RBM, neurons are random, which only have active and inactive states, like BM, RBM also have hidden layers and visible layers. The difference is RBM has no connections inside the layers.
Deep Belief Networks (DBN)
DBN is constructed by multiple layer RBM, every low level RBM output as the input of higher level RBM. Comparing with traditional BP, DBN can get suitable initial weight of nets by pre-training, which can conquer the localized optimization problem of BP caused by initialization.
This complete training is achieved in two steps:
* Unsupervised training on each RBM layer, making sure as many as possible features are preserved from lower to higher layers.
* Fine adjust the DBN weight by BP, after getting forward and backward weights.
DBN defects detection analysis:
* Solar cell surface have simple pattern, the defect-free area have uniform pattern as same background and detection location, while the defected image only have difference on grey scale.
* DBN can combine processing image and extracting features, which is fast for production lines.
* DBN can be combined with traditional detection nets
Process and results
Images are normalized to 64 x 64 and 100 images are used for training while 20 for testing. The defects present include: fractures, scratches, missing corners.
The network is then trained with the previously explained process yielding a total training time of 2597s. After training image reconstruction can be done using the 20 images. The test set takes 3.01s, averaging 0.151s per frame. In practice, this method still has shortcomings such as, using high resolution images which directly impacts training time.
If you want to know more about DBN’s check http://videolectures.net/mlss09uk_hinton_dbn/
