Automatic Reading Systems (OCR)
Optical character recognition or OCR is a technology in full expansion that includes many image processing techniques like segmentation, classification and pattern recognition. This field of artificial intelligence enables us to develop automatic reading systems, which aim to increase treatment speed of paper documents but also other perspectives like licence plates for intelligent visual surveillance.
Scientific topic being studied for a long time, OCR allows dealing with various kinds of text documents. Indeed, numerous constraints such as colours, different character fonts and sizes, bad document orientation or lighting and the use of different acquisition devices (scanners or digital cameras) can all operate. To date, there is no generic OCR. Certain techniques are necessary in order to have optimal results for a specific application.
Many methods are developed at Multitel for every steps of the automatic reading system:
- Text localisation: the text zones in an image need to be located (separation of figures and lines into forms, text localisation in outdoor
Automatic text localisation system
- Text segmentation in distinct character units: methods are based on statistical analysis (histograms, local variances) and on connected components.
- Characters recognition: various classifiers can be employed.
Technologies of statistical pattern classification can be classified in three
categories: parametric methods (hypotheses about parameters distribution),
non-parametric methods (comparison of unknown objects and reference objects) and
artificial neural networks. Our researches focus on artificial neural networks
with multiple layers. They have the advantage of offering comparable
performances with more moderate computational requirements.
Symbolic of artificial neural networks
Many works are done in the field for handwritten or printed characters in poor quality images. Efficient methods of feature extraction were developed in order to reach excellent recognition rates. Our error rate is less than 5% for the 26 capital letters and for the 10 digits in the context of multi-writers.
More robust hybrid systems like combination of neural networks and statistical errors correction based on semantic and graphological analysis, can improve the performances.
Application fields of text recognition include: document conversion into digital information that will facilitate and improve a company’s management, archiving and sharing of information; automatic sorting of mail or labelled products as well as printing control. These applications can extend speech synthesis’ possibilities and help partially-sighted people in improving their daily life. This application is currently being developed.
Medical imaging is an essential tool for most radiologists. Diagnosis is a time-consuming process that presents an inherent variability. By using advanced image processing techniques and automatic methods, we are able to reduce these drawbacks by increasing accuracy and efficiency.
At Multitel, we are working on such projects. The experience that we have acquired while developing imaging solutions allows us to be an active player in the market place.
The first research topic is about automatic detection of pulmonary emboli. Research is also performed in the field of intra- and post-operative follow-up for surgery and radiotherapy of the neck and the brain.
- Automatic Detection of Emboli :
Detection of emboli is divided in two parts: segmentation of the pulmonary arterial tree and its
Given the tubular, hierarchical structure of the pulmonary arteries, the segmentation requires a particular method (such as region growing, fast marching and level set) for 3D images.
Anatomical knowledge models may be associated to increase the robustness of the method. Analysis can then be performed through, in particular, local histograms and concavity
Pulmonary arteries segmentation
- Intra-operative Follow-up :
Initial planning of surgical operation is carried out after the segmentation of a tumour. This is a partially automatic process. Following an image pre-processing, this algorithm finds the tumours by providing three distinct areas: inside, outside and fuzzy area. This segmentation can be performed on various modalities, each bringing its specific information which is often complementary.
For the follow-up, original images are co-registered on intra- or post-operative ones using similarity criteria through rigid or non-rigid transformations.
Co-registration of images allows a measure of the evolution of the tumour during its treatment and further adapts it.
Research works in close collaboration with hospitals whose radiologists are the eventual users of these tools. This collaboration allows Multitel to have access to the images and ensures that the developments correspond to the needs.
Automatic Quality Sorting of Apples
An inspection system is basically composed of a camera, an acquisition system and a software. Pictures of the material to be inspected are taken by the camera and sent to the computer by the acquisition system, where the software analyses them and decides.
Current sorting systems in industry food processing are generally based on human work. This work process could be automated in order to increase the processing speed as well as to reduce the error introduced by humans.
The European Union defines strict standards based on quality, shape and colour for sorting agricultural products, which make the sorting task even harder. Within the agricultural products, apples are probably the hardest ones to be sorted by machine vision due to their widely changing skin colours and the various types of defects they are exposed to.
There is an ongoing project for ‘Automatic quality classification of apples’, which has the main goal of classifying apples by image processing and artificial neural network techniques at a rate of 10 apples/sec or more.
In short, the computer processes the images from different views and decides which quality class the apple should belong to.
Automatic Segmentation of Defects on ‘Jonagold’ apples