The potential of deep learning in the pharmaceutical market
Deep Learning is a technology that is based on the structure of the human brain. It is increasingly implemented in industrial image processing – and is now very often used to extend and complement rule-based image processing. With Deep Learning methods, the neural network learns to reliably detect anomalies by means of example images. Typical applications include the detection of cosmetic defects such as scratches, stains and dirt. The potential for our machines - especially for the manufacture of pharmaceutical and medical products - is obvious.
What are the differences to rule-based image processing?
In traditional rule-based image processing, one or more objects in the image are contrasted and isolated using edge finders or thresholding methods to check for quality. Applications such as the detection of anomalies with a low contrast to the background are very challenging for rule-based image processing and the programming effort is enormous. With Deep Learning and Convolutional Neural Networks (CNNs), we use the already mentioned example images to train a neural network specifically for such applications. The training process involves methods such as Supervised and Unsupervised Learning."
Will rule-based image processing become superfluous?
Rule-based image processing will continue to play an important role in the future too. Some tasks are difficult or impossible to solve with Deep Learning. These include the precise measurement of objects or the identification of barcodes and data matrix codes. At Harro Höfliger, we therefore do not consider Deep Learning-based image processing as a replacement, but as a valuable complement to rule-based image analysis. It is precisely the interaction of both methods that ensures an ideal and process-safe quality control. The combination of both methods not only contributes to precise production, but also to the better and faster detection of product defects. To this end, we rely, among other things, on the potential of smart cameras and PC-based camera systems.
What role does hardware play?
Indeed, Deep Learning places comparatively high demands on hardware. Especially training the neural networks with the high amount of data and training cycles is extremely CPU-intensive. In contrast to conventional CPUs, which very quickly reach their performance limits for such tasks, modern GPUs (graphics processors) are used. Due to the large number of cores, a high degree of parallelism can be achieved, which significantly accelerates both training time and inferencing. The graphic card's high power consumption of 200 watts and more is achieved by using a correspondingly efficient power supply. Only by using well-coordinated and adapted hardware, Deep Learning image processing systems can be implemented even at high machine speeds and cycle rates.
Can you give an example of a current project?
We are currently working on a solution for the position and alignment check of tubular bags on a conveyor belt, by means of a picker cell. For this, the plastic ampoules in the bag must be inserted into the box with the closure side facing upwards. Despite the non-transparent aluminum foil bag, the position of the flow packs can be easily determined with today’s image processing systems, based on X-Y coordinates and rotation. The situation is somewhat different when it comes to the alignment of the bags on the belt. Here, only minimal information such as the sealing contour can give information on the plastic ampoules’ position in the bag. Thanks to Deep Learning and its classification methods for objects we were able to provide our customer with a good solution for this challenging task.