Abstract. Defects detection on images is a current task in quality control and is often integrated in partially or fully automated systems. Assessing the performances of defects detection algorithms is thus of great interest. However, being application and context dependent, it remains a difficult task. This paper describes a methodology to measure the performances of such algorithms on large size images in a semi-automated defect inspection situation. Considering standard problems occurring on real cases, a comparison of typical performance evaluation methods is made. This analysis leads to the construction of a simple and practical ROC-based method. This method extends the pixel-level ROC analysis to an object-based approach by dilating the ground-truth and the set of detected pixels before calculating true positive and false positive rates. These dilations are computed thanks to the a priori knowledge of a human defined ground-truth and gives to true positive and false positive rates more consistent values in the semi-automated inspection context. Moreover, dilation process is designed to be automatically suited to the objects shape in order to be applied on all types of defects.

No abstract

Sextant Avionique and Thomson LCD developped a 6"x8" portrait AMLCD for avionic cockpit display applications. Main features of the display are : 504x672 resolution in RGGB quad, wide viewing angle (± ±60°) and low reflectivity (0.8%). In addition, the mechanical design of the display module allows low maintenance costs and withstanding of harsh environmental conditions.

We present in this letter a noniterative learning rule for classification and neural networks, which allows to eliminate the drawback of overfitting of the pseudo-inverse (PI) solution and to preserve good learning performances. This solution, which is obtained by artificially increasing the number of patterns in the learning set, is a parametric form between the pseudo-inverse and the Hebb solutions. The results are compared to each other and with those of a gradient descent iterative procedure on two very different examples. We show that the proposed solution is near to the one of the iterative procedure.

Defect detection in images is a current task in quality control and is often integrated in partially or fully automated systems. Assessing the performances of defect detection algorithms is thus of great interest. However, because this is application-and context-dependent, it remains a difficult task. We describe a methodology to measure the performances of such algorithms on large images in a semi-automated defect inspection situation. Considering standard problems occurring in real cases, we compare typical performance evaluation methods. This analysis leads to the construction of a simple and practical receiver operating characteristic (ROC) based method. This method extends the pixel-level ROC analysis to an object-based approach by dilating the ground truth and the set of detected pixels before calculating the true-positive and false-positive rates. These dilations are computed thanks to the a priori knowledge of a human-defined ground truth and gives to true-positive and false-positive rates more consistent values in the semi-automated inspection context. Moreover, the dilation process is designed to be automatically suited to the object's shape in order to be applied on all types of defects without any parameter to be tuned.

In order to assess the quality of an X-ray imager it is necessary to measure the visibility of any artifact that might be present in the image. Several methods have been proposed in the literature to calculate this visibility. To predict the performance of these methods in the context of quality control of X-ray imagers, a base of 10 artifacts as different as possible in shape and aspect have been created (a pixel, a line, a step, various spots and noises). The amplitude for which each artifact has a probability of 50% to be detected has been determined. To do so, the artifacts have been observed merged with three noises of different spectra ("white noise", "high-frequency" noise and "low-frequency" noise). To determine the 50% detection probability amplitudes, a variant of the 2 Alternative Forced Choice procedure has been used. It has been checked that the measurement exploitation method is unbiased and its precision is sufficient. The dispersion of results between testers, around 15% in average, is also satisfactory. These results are a solid and objective basis to check the relevance and limits of visibility measurement methods described in literature, applied to the domain of quality control of X-ray imagers.

No abstract

In medical imaging, model observers such as the "Hotelling observer" and the "Non Prewhitening Matched Filter" have been proposed to detect objects in X-ray images. These models, based on decision theory, are applied over the entire image. In this paper, we developed a model that mimics some processes of human visual perception. The proposed model is locally applied on some particular areas that correspond to the salient areas of the object. By doing this, the model mimics the sequence of eye fixations that we make when we explore an image for example in order to detect an object. The study is divided into three parts: a psychophysical experiment to obtain human's performance to detect various objects in noises, a theoretical part to develop the proposed model, and finally, a result part. During the experiment, several participants were asked to detect objects in noisy images using a free search task. The luminance contrast of objects was adaptively adjusted according to their responses to obtain a percentage of correct detection for each object of 50 %. The proposed model, based on decision theory, was applied locally on some areas of the image that has a size corresponding to the high visual acuity of foveal vision. Areas were chosen according to their high saliency values computed through a bio-inspired model of visual attention. For each area, our model returned a detectability index. By supposing statistical independence between areas, the local indexes are combined into a global detectability index. Results show that the proposed model fits the results of the psychophysical experiment and outperforms classical models of the literature.