deep-learning - Incremental learning for image similarity (Siamese Network) and predicting the label

Question

The requirement of the task is to predict correct labels for the correct product. Suppose, a company produces 100 cork disks with a serial number. Then it captures the surface images of cork disks and trains a neural network (Siamese/CNN/Embeddings) model with the augmented part of the same image for training the texture similarity invariant to rotation, translation or change in brightness. Now the products are shipped to the destination along with this trained model but the images are discarded and new images are captured that have arrived. When these new images (which are different in brightness, translation and rotation from the original images captured at the company and on which the model was trained) are fed into the model, it should predict the correct serial number corresponding to the products.

I have already trained a Siamese network which gives good similarity score when 2 images are compared (one from the company and one taken at the destination). But there are 2 problems with this approach:

1. During testing, the old images won't be available. They will be discarded since the images are high resolution and takes a lot of space. So we need to train a model which properly maps images to some embedding vectors that we can store and maps it to the serial number.

2. The model needs to be trained incrementally. Another 1000 disks will be produced the next day and the model should be tuned according to the new dataset.

Any help on this topic would be really appreciated!

标签： deep-learningcomputer-visionconv-neural-networksiamese-network