Machines for pressing pills in the pharmaceutical industry eject more than one hundred pills per second. Defective pressed pills are difficult to identify. Reply has developed a solution to this problem for a pharmaceutical company that uses two high-speed cameras to capture the front and backs of the pills.
# Image Recognition
# Deep Learning
# Transfer Learning
Machines for pressing pills in the pharmaceutical industry eject more than one hundred pills per second. Defective pressed pills are difficult to identify with the conventional methods of pill inspection devices, as the defects are often very small and hardly detectable. The long reaction times of the testing procedures lead to large quantities of defective pills that subsequently have to be sorted out.
Reply has developed a solution to this problem for a pharmaceutical company that uses two high-speed cameras to capture the front and backs of the pills. A deep convolutional auto-encoder neural network analyses each image for deviations, reports the results to a dashboard and, if necessary, stops the production process. This transfer learning process has made reliable detection accuracy possible even for very small defects. The result is short response times, early defect detection, and meaningful error messages to eliminate the problem. This automatic inspection of every single pill leaving the press ensures that the defective pills are sorted out before they are packaged. The number of defective pills produced can be reduced and is also easily adaptable to different types and shapes of pills.
Why is Deep Learning the way to go? Due to increased computing power, it has become possible to have neural networks perform image analyses, which has several advantages. First of all, a Deep Learning algorithm automatically extracts features for the user. In the past, you needed an expert to develop such features. That was very time-consuming, as the manually created features had to be invariant to rotation, for example. Deep Learning is also very robust to changes in the image. For example, if you have classified an object in an image as a dog, you want the model to recognise a dog regardless of its size or colour, and regardless of whether it is sitting upright or lying in a tree. The same is true for our pills. Once the network has learned the correct features, you are independent of turns or different lighting conditions. In general, features are more robust in Deep Learning, which also makes them universally reusable.
This means that you can learn features by training your network on a general dataset and then reusing it for different classification tasks, which is called transfer learning. Although training takes a while, the Deep Learning network can be quickly and easily applied to a large number of images afterwards, rather than having to be computed from scratch for each image. So, it is more of a greedy algorithm: once you know the weighting, forward propagation is just a quick linear transformation.
Deep Learning techniques can thus be used in the pharmaceutical industry, for example, to detect even the smallest deviations from the correct pill shape. The results and deviations can be evaluated in real time on a dashboard application. The high-speed detection of broken pills enables production to be stopped at an early stage, thus helping to ensure quality and avoid unnecessary troubleshooting costs.