Machine learning in visual inspection

Machine learning methods are used to tackle a wide range of visual inspection challenges. Data based methods can distinguish different spectra of materials from each other. Collecting a high number of data points to train a suitable classifier can solve challenging problems. Close work with our customers enables to consider knowledge of specific domains for a certain task. If required, we cooperate with other Fraunhofer Institutes for tasks in classification of wood, plastics, minerals, food or tobacco.

Foreign substances, which are very similar in color to the tobacco itself, have to be sorted out during the tobacco sorting process. The usage of a RGBN (RGB + Near infrared-channel) camera and classical machine learning algorithms enable the detection of foreign materials. Information from four channels are reduced to three by dimension reduction. As a result, an image with false colors is created, which shows a fusion of the reduced information. So it is easier to handle and advantageous for real-time sorting.

In another pretest, the separability of different soft- and hardwoods was investigated. For example, fine chips of pine should be separated from chips of beech and birch by recording of a hyper-spectral short wave infrared camera. Good results were achieved by a non-linear classification afterwards.

We offer a complete pipeline for visual inspection: From correct preselection of the measurement system and data acquisition to suitable selection of a classifier and real-time implementation in the industry. Our competences in optical data acquisition and AI go well together.