Machine learning (deep learning) and neural networks

Deep learning methods, sometimes also reffered to as neural networks, which are a part of a broader category of algorithms known as machine learning, are able to learn from sample image data how to properly evaluate the data. Instead of inventing and programming complex rules, a special training algorithm is able to “learn” these rules and dependencies on its own. This is done by presenting (usually many) sample images to the training algorithm, along with the correct results. The output of the training is then a so-called neural network model, which is used to evaluate all other (i.e., even the model’s “previously unknown”) images.

While deep learning methods can greatly simplify the solution of many more complex machine vision tasks, their proper deployment also requires some expertise. Depending on the type of task, the type of model and possibly also the architecture of the neural network must be chosen correctly. Optimal model parameter settings, training methodology and correct annotation of sample images are also important. As with all machine vision tasks, the design of the correct image capture system (selection and optimal configuration of camera, optics and lighting) also plays a primary role. All these factors contribute significantly to the success of solving a machine vision task using deep learning methods.

Thanks to the fact that our experts have been working in the field of deep learning and neural networks for a long time and have tested their skills on many projects and studies, we can cope with very non-standard customer requirements.

Surface inspection and the detection of bumps, defects, but also scratches or dirt is one of the typical tasks for deep learning and neural network methods. Once a sufficiently large dataset is loaded, the neural network is able to teach itself to detect a variety of defects and material flaws that match the task.

• It can also be used in problems with hard-to-define product defects where conventional methods fail.
• With sufficient training data, machine learning can often achieve higher evaluation accuracy – i.e., when checking for defects, there is a lower rate of both undetected defects and false detections.
• Universal adaptation to various changes in production (changes in material, product type, etc.) – unlike conventional methods, there is a straightforward and relatively fast solution by retraining the neural network model.
• Enhanced defect analysis capabilities – deep learning provides greater capabilities not only in the detection of individual defects, but also in their classification (accurate defect classification is often not possible using conventional methods).
• A uniform approach for different types of products and different types of controls on products.
• The model can be taught continuously with production, the change of the learned model is solved by loading one model file – it can be done even in full operation.
• Advanced methods can be used to detect defects that are difficult to detect by eye, such as cracks on solar panels, chips, wafers, etc.
• By varying the position of the object, it allows you to select the region of interest (ROI) even if its boundaries are blurred, partially overlapped, or complexly defined, this allows you to focus only on the area of the image of interest each time and thus significantly refine its evaluation.

• The process of training models is random, so no two models can ever be trained for the same training data (However, the image evaluation itself is already strictly deterministic).
• Deep learning cannot be applied to evaluations based on precise measurements of the dimensions and angles of a given object; conventional image processing methods must be used here.
• To successfully train the model, it is necessary to have a certain number of images for each category (the so-called dataset) – for example, with conventional methods we are able to design an algorithm to detect a certain product defect even if there is only one sample of this defect. In deep learning this is not possible, an adequate number of pieces of each defect must be available (usually at least around 10 pieces).
• A neural network model is a “black box” that you cannot look into and see why it evaluated a given image with a particular result and not otherwise. There are, however, visualization tools – called probability maps – that provide at least partial insight into the model’s “decision-making” process.
• Slightly higher price for special software applying deep learning methods – this is relative, however, because when evaluating with conventional methods there are higher costs for the development of image processing algorithms, or some tasks are not solvable at all.