6533b7d4fe1ef96bd1261a46

RESEARCH PRODUCT

Apparence des matériaux, Vision artificielle, Inspection qualité, Reflectance Transformation Imaging

subject

description

Mastering the visual perception of the surfaces of manufactured products is a central issue for industry. However, in industry, the quality of surfaces is often assessed by human inspectors. Only a few specific cases use an instrumental or photometric approach. Among the photometric approaches, one of them is experiencing significant growth: Reflectance Transformation Imaging (RTI). However, this technique has limitations in terms of data acquisition and processing. The objective is therefore to correct some of these limitations in order to improve the RTI and, consequently, the visual quality control of surface conditions in industry.The current RTI systems are limited and cannot meet our needs in terms of implementation and experimentation of the modalities and methods related to RTI. We therefore developed an RTI measurement system coupled with a control software. This set allows us access to hardware and software code to add, modify, and control the parameters and acquisition methods. One of the developments consisted in implementing a new modality of acquisition which consists in coupling the High Dynamic Range (HDR) to the RTI (HD-RTI). This coupling makes it possible to correct a measurement bias linked to the exposure time of the camera and to the limit of the sensor in terms of dynamic range. The HD-RTI measure the full dynamic range of the luminance response of the inspected surfaces. With HD-RTI stereo-photometric data, we can virtually reconstruct the scene by simulating an arbitrary exposure time, but also better characterize and therefore discriminate surface anomalies.RTI generates large amounts of data, which becomes more complex depending on the acquisition methods used, such as HD-RTI. We propose a methodology to characterize the appearance of surfaces from RTI measurements. The characterization of surface states is based on the use of geometry descriptors and the photometric behavior of surfaces. The variety of descriptors allows a fine characterization of the different surface states. From the descriptors extracted, from the RTI acquisitions, we propose a method to estimate the multi-scale and multi-level visual saliency in each pixel and thus make it possible to discriminate surface anomalies. A methodology, to segment RTI data using saliency, and determine the most relevant descriptors to use according to a global criterion, is then applied to an application case. Then, distance calculation is extended to RTI acquisitions in order to compare surface states. The distance is correlated with the degree of difference between the characteristics of the surface finishes. Finally, a distance is also calculated between the appearance reconstruction models.