Apples sorting by using microwave imaging

Nondestructive evaluation for agricultural products is of increasing interest. The researchers investigate the use of microwave imaging systems, which allow non-destructive control of fruits while monitoring their growth and maturity. As for every fruit, the apples face many diseases, which is an economic and an agricultural issue. The objective of this work is to develop a non destructive and an automatic method to sort damaged fruits from the healthy ones by microwave imaging. This work is divided into six chapters, in chapter 1, we begin with a state of the art of the destructive and non destructive methods which are used in quality control of the agricultural products. In chapter 2, we defined the classification methods which are used in this work and also, we described the optimization methods which are used for finding an optimal values of hyper-parameters of classification method. In chapter 3, a non-destructive and an automatic method to sort damaged fruits from the healthy ones by using microwave imaging is presented . The proposed method is a proof of concept that starts with millimeter wave measurements in W-band, which are pre-processed by the two-dimensional Fourier transform (2D-FFT) and segmentation methods, and ends with the classification using Machine Learning algorithm. This method has been validated first, on the apples with a classification error of 6% and then on the peaches where the classification error of 0%. In chapter 4, we addressed the impact of frequency and spatial diversity on the accuracy of the classification, indeed, this study highlights that the limitations of the proposed method are consistent with the physical phenomena. The first results showed the potential of microwave imaging and artificial intelligence to sort fruits, but the acquisition time remains prohibitive for an industrial application, so two investigations were conducted, the first one concerns the simplification of the acquisition scheme by measuring the amplitude of the electromagnetic field. In addition, the preprocessing is also simplified, by removing the 2D-FFT. The second study concerns the reduction of spatial diversity by optimizing the number of measurement points, which automatically leads to a reduction of the measurement time. Using the Grey Wolf Optimizer, the number of measurement points is reduced from 22801 to 196.In chapter 5, the proposed method has been adapted and tested on archaeological samples dating back to more than 8000 years before our era, in order to trace in time and space the introduction of farming in Europe. We aim at identifying two types of pottery shards: "spirals" and "coiling". A new method is developed, which is similar to the one used for fruits but with a new preprocessing step and the use of a neural network of MLP type. The accuracy of the classification is of 99%. As for the fruits, the measurement system was optimized by elaborating two scenarios and using one of the bio-inspired optimization methods developed during this thesis and tested for the first time on real data. The classification error of both optimized systems is less than 1%. At chapter 6, a new approach for the non-destructive and automatic sorting of fruit has been studied, it complies with the requirements of the industry regarding the acquisition time because it requires only one antenna operating on a wide-band. We elaborated two strategies to detect the fruit defect. In the first one, the fruits are sorted in the frequency domain by using the Short Time Fourier Transform method. The second strategy concerns the fruits sorting in the time domain using the inverse of the Short Time Fourier Transform method. For both strategies, the classification error is 1%. The advantages and disadvantages of this new approach are discussed.