Vincent Leemans

A real-time grading method of apples based on features extracted from defects

This paper presents a hierarchical grading method applied to Jonagold apples. Several images covering the whole surface of the fruits were acquired thanks to a prototype grading machine. These images were then segmented and the features of the defects were extracted. During a learning procedure, the objects were classified into clusters by k-mean clustering. The classification probabilities of the objects were summarised and on this basis the fruits were graded using quadratic discriminant analysis. The fruits were correctly graded with a rate of 73%. The errors were found having origins in the segmentation of the defects or for a particular wound, in a confusion with the calyx end.

Defects segmentation on ‘Golden Delicious’ apples by using colour machine vision

Abstract A method based on colour information is proposed to detect defects on ‘Golden Delicious’ apples. In a first step, a colour model based on the variability of the normal colour is described. To segment the defects, each pixel of an apple image is compared with the model. If it matches the pixel, it is considered as belonging to healthy tissue, otherwise as a defect. Two other steps refine the segmentation, using either parameters computed on the whole fruit, or values computed locally. Some results are shown and discussed. The algorithm is able to segment a wide range of defects.

On-line Fruit Grading According to their External Quality using Machine Vision

This paper presents apple grading into four classes according to European standards. Two varieties were tested: Golden Delicious and Jonagold. The image database included more than a 1000 images of fruits (528 Golden Delicious, 642 Jonagold) belonging to the three acceptable categories—Extra, I and II—and the reject (each class represents, respectively, about 60, 10 and 20% of the sample size). The image grading was achieved in six steps: image acquisition; ground colour classification; defect segmentation; calyx and stem recognition; defects characterisation and finally the fruit classification into quality classes. The proposed method for apple external quality grading showed correct classification rates of 78 and 72%, for Golden Delicious and Jonagold apples, respectively. Taking into account that the healthy fruit were far better graded and considering that this class was under represented in the sample compared with the fruit population, the results of the proposed method (an error rate which drops to 5 and 10%, respectively) are compatible with the requirements of European standards.

Defect segmentation on 'Jonagold' apples using colour vision and a Bayesian classification method

This paper shows how the information enclosed in a colour image of a bi-colour apple can be used to segment defects. A method to segment pixels, based on a Bayesian classification process, is proposed. The colour frequency distributions of the healthy tissue and of the defects were used to estimate the probability distribution of each class. The results showed that most defects, namely bitter pit, fungi attack, scar tissue, frost damages, bruises, insect attack and scab, are segmented. However, russet was sometimes confused with the transition area between ground and blush colour.

Line cluster detection using a variant of the Hough transform for culture row localisation

An adaptation of the Hough transform was proposed for the detection of line clusters of known geometry. This method was applied in agriculture for the detection of sowing furrows created by a driller and of chicory plant rows during harvesting process. The sowing rows were revealed by a background correction, the background being obtained thanks to a median rank filter. The method was found efficient in eliminating the shadows. For the crop rows, a neural network was used to localise the plants. While the petiole and the leaves were easily separated from the soil, the chicory root and the soil having about the same colour and the lighting condition varying widely, it was more difficult to obtain a good contrast between those parts, which leaves place for some improvements. The adapted Hough transform consisted in computing one transform for each line in the cluster with, for reference, the position and direction of the theoretical position of the row. The different transforms were then added. It was found effective for both the sowing rows and the chicory rows. Results remained good even in very noisy conditions, when the rows were incomplete or when artefacts would lead its classical counter part to show several alignments other than the expected ones. The culture rows were localised with a precision of a few centimetres, which was compatible with the proposed applications.

AE—Automation and Emerging Technologies: On-line Fruit Grading according to their External Quality using Machine Vision

This paper presents apple grading into four classes according to European standards. Two varieties were tested: Golden Delicious and Jonagold. The image database included more than a 1000 images of fruits (528 Golden Delicious, 642 Jonagold) belonging to the three acceptable categories—Extra, I and II—and the reject (each class represents, respectively, about 60, 10 and 20% of the sample size). The image grading was achieved in six steps: image acquisition; ground colour classification; defect segmentation; calyx and stem recognition; defects characterisation and finally the fruit classification into quality classes. The proposed method for apple external quality grading showed correct classification rates of 78 and 72%, for Golden Delicious and Jonagold apples, respectively. Taking into account that the healthy fruit were far better graded and considering that this class was under represented in the sample compared with the fruit population, the results of the proposed method (an error rate which drops to 5 and 10%, respectively) are compatible with the requirements of European standards.

A sorting optimization curve with quality and yield requirements

Binary classifiers used for sorting can be compared and optimized using receiver-operating characteristic (ROC) curves which describe the trade-off between the false positive rate and true positive rate of the classifiers. This approach is well suited for the diagnosis of human diseases where individual costs of misclassification are of great concern. While it can be applied to the sorting of merchandise or other materials, the variables described by the ROC curve and its existing alternatives are less relevant for that range of applications and another approach is needed. In this paper, quality and yield factors are introduced into a sorting optimization curve (SOC) for the choice of the operating point of the classifier, associated with the prediction of output quantity and quality. Given examples are the sorting of seeds and apples with specific requirements. In both cases the operating point of the classifier is easily chosen on the SOC, while the output characteristics of the sorted product are accurately predicted.

Yield variability linked to climate uncertainty and nitrogen fertilisation

At the parcel scale, crop models such as STICS are powerful tools to study the effects of variable inputs such as management practices (e.g. nitrogen (N) fertilisation). In combination with a weather generator, we built up a general methodology that allows studying the yield variability linked to climate uncertainty, in order to assess the best N practice. Our study highlighted that, applying the Belgian farmer current N practice (60-60-60 kg N/ha), the yield distribution was found to be very asymmetric with a skewness of -1.02 and a difference of 5% between the mean (10.5 t/ha) and the median (11.05 t/ha) of the distribution. This implies that, under such practice, the probability for farmers to achieve decent yields, in comparison to the mean of the distribution, was the highest.

Assessment of plant leaf area measurement by using stereo-vision

The aim of this study is to develop an alternative measurement for the leaf area index (LAI), an important agronomic parameter for plant growth assessment. A 3D stereo-vision technique was developed to measure both leaf area and corresponding ground area. The leaf area was based on pixel related measurements while the ground area was based on the mean distance from the leaves to the camera. Laboratory and field experiments were undertaken to estimate the accuracy and the precision of the technique. Result showed that, though the leaves-camera distance had to be estimated precisely in order to have accurate measurement, the precision of the LAI evaluation, after regression, was equivalent to the reference measurements, that is to say around 10% of the estimated value. This shows the potential of the 3D measurements compared with tedious reference measurements.

Estimation of leaf nitrogen concentration on winter wheat by multispectral imaging

Precision agriculture can be considered as one of the solutions to optimize agricultural practice such as nitrogen fertilization. Nitrogen deficiency is a major limitation to crop production worldwide whereas excess leads to environmental pollution. In this context, some devices were developed as reflectance spot sensors for on-the-go applications to detect leaves nitrogen concentration deduced from chlorophyll concentration. However, such measurements suffer from interferences with the crop growth stage and the water content of plants. The aim of this contribution is to evaluate the nitrogen status in winter wheat by using multispectral imaging. The proposed system is composed of a CMOS camera and a set of filters ranged from 450 nm to 950 nm and mounted on a wheel which moves due to a stepper motor. To avoid the natural irradiance variability, a white reference is used to adjust the integration time. The segmentation of Photosynthetically Active Leaves is performed by using Bayes theorem to extract their mean reflectance. In order to introduce information related to the canopy architecture, i.e. the crop growth stage, textural attributes are also extracted from raw images at different wavelength ranges. Nc was estimated by partial least squares regression (R² = 0.94). The best attribute was homogeneity extracted from the gray level co-occurrence matrix (R² = 0.91). In order to select in limited number of filters, best subset selection was performed. Nc could be estimated by four filters (450 ± 40 nm, 500 ± 20 nm, 650 ± 40 nm, 800 ± 50 nm) (R² = 0.91).