Neural network for bulk sorting

A bulk sorting system for sorting objects in bulk into at least one set of wanted objects and at least one set of unwanted objects, the bulk sorting system comprising:

• at least one radiation source arranged to radiate the objects received by the system in bulk;
• at least one optical sensor arranged to capture reflected radiation of the objects received by the system in bulk and acquire the reflected radiation as multi- or hyperspectral data, said at least one optical sensor comprising a near-infrared (NIR) scanner arranged to scan the objects received by the system in bulk, wherein the multi- or hyperspectral data comprises the scan data, wherein the multi- or hyperspectral data is a multi-dimensional image where each pixel or equivalent super pixel or grouping of pixels comprises a number of bands with order of magnitude 1-10 with narrow or wide spectrums with order of magnitude 100-1000 nm or a multi-dimensional image where each pixel or equivalent super pixel or grouping of pixels comprises a number of bands with order of magnitude 4-500 with narrow spectrums with order of magnitude 1-50 nm;
• a processing circuit configured to analyze the reflected radiation of the objects by inputting the multi- or hyperspectral data into a convolutional neural network (CNN) with at least two convolutional layers in order to either detect and classify the objects in the multi- or hyperspectral data and/or semantically segment the multi- or hyperspectral data;
• and
• said CNN using as an input the scan data in order to either detect and classify and/or semantically segment objects in the scan data as said at least one set of wanted objects and said at least one set of unwanted objects,
• wherein in the instances where objects received in bulk by the system are overlapping and/or stacked with other objects, the CNN identifies objects in the scan data of said overlapping objects as a single group output,
• a mechanical sorter configured to sort the objects according to their classification and/or segmentation using the analysis of the processing circuit including the single group output such that different overlapping and/or stacked objects, as analyzed by the processing circuit, are treated as a single group of objects.