Tri de pommes par imagerie millimétrique

This work is a proof of concept proposing and describing a complete workflow to differentiate healthy from damaged apples, staring from mm-W measurements and ending-up with classification based on Support Vector Machine. The method has proven to be successful with only 6% error when scan angle and frequency diversity are used. In a first step, we build a database of more than 1800 images obtained by processing measurements with a 2D-FFT. Images are then binarised and used as input of a nonlinear SVM. At this stage, 90% of the data base is used for training and coefficients C and gamma are tuned to reach the smallest error. The 10% remaining images are used for the test. In a second step, we assess and discuss the influence of the physical inputs of the database : the frequency, the sparsity in measurements points and the size of the apples. Finally we explore new scenarios considering other fruits.