Apparatus for sorting raw coffee beans based on defect detection
A bean defect sorter includes a container configured to store beans, a transfer guide configured to receive the beans from the container and to guide each bean in series, the transfer guide being configured to discharge the beans one by one, a rotary configured to rotate in a predetermined direction and to receive each of the beans discharged from the bean dispenser, an imaging device configured to provide a bean image of one or more of the beans, a processor including a learning algorithm that is configured to classify each bean as one of a normal bean, a defective bean, and an indeterminable bean, a bean sorter configured to discharge each bean from the rotary by using air pressure according to a classification result, a storage configured to store each of the normal bean, the defective bean, and the indeterminable bean, and a residual bean collector to process escaped beans.
1 . A bean defect sorter comprising:
an imaging device configured to capture images of beans;
a bean rotary compartment configured to rotate and position each bean for imaging by the imaging device;
a processor including a learning algorithm that is configured to:
receive the images from the imaging device,
determine, from the images, whether each bean has a defect, and
classify each bean as one of a normal bean, a defective bean, and an indeterminable bean based on image features extracted by the learning algorithm;
a bean sorter configured to physically separate each bean from the bean rotary compartment according to a classification result from the processor; and
a sorted bean storage configured to store each of the normal bean, the defective bean, and the indeterminable bean in separate storage spaces.
2 . The bean defect sorter of claim 1 , wherein the learning algorithm comprises a Convolutional Neural Network (CNN).
3 . The bean defect sorter of claim 2 , wherein the learning algorithm further comprises a Residual Network (ResNet).
4 . The bean defect sorter of claim 1 , wherein the imaging device comprises:
a first camera configured to capture images of a top surface of each bean on the bean rotary compartment; and
a second camera configured to capture images of a bottom surface of each bean, and
wherein the processor is configured to merge the images from the first camera and the second camera to provide a complete image of each bean for classification.
5 . The bean defect sorter of claim 1 , wherein the processor is configured to divide pre-training images into training and validation.
6 . The bean defect sorter of claim 1 , wherein the processor is configured to:
perform data augmentation on training images used to train the learning algorithm, the data augmentation comprising applying variations to the training images to increase diversity and quantity of the training images.
7 . The bean defect sorter of claim 6 , wherein the processor is configured to:
perform normalization and pixel mean subtraction on the training images; and
set hyper-parameter values including a batch size, an epoch size, and a learning rate for training the learning algorithm.
8 . The bean defect sorter of claim 1 , wherein the bean rotary compartment defines a plurality of seating spaces, each of the plurality of seating space being configured to receive one of the beans guided to the bean rotary compartment.
9 . The bean defect sorter of claim 8 , wherein the bean sorter comprises:
a plurality of air pumps; and
a plurality of air tubes connected to the plurality of air pumps, respectively, each of the plurality of air tubes being configured to supply air from one of the plurality of air pumps toward one of the plurality of seating spaces of the bean rotary compartment to thereby physically separate each bean from the bean rotary compartment according to the classification result.
10 . The bean defect sorter of claim 9 , wherein the bean rotary compartment further defines a plurality of bean-drop openings that are in fluid communication with the plurality of seating spaces, respectively, and configured to provide a passage for sorted beans to the sorted bean storage.
11 . The bean defect sorter of claim 1 , wherein the bean rotary compartment comprises:
at least one disk plate that defines a plurality of seating spaces, each of the plurality of seating spaces being configured to receive one of the beans; and
a support plate that is made of a transparent material and mounted to the at least one disk plate, the support plate defining bottom surfaces of the plurality of seating spaces configured to seat the beans thereon.
12 . The bean defect sorter of claim 11 , wherein the imaging device comprises:
a first camera positioned above the support plate and configured to capture images of a top surface of each bean on the support plate; and
a second camera positioned below the support plate and configured to capture images of a bottom surface of each bean through the support plate, and
wherein the processor is configured to merge the images from the first camera and the second camera to provide a complete image of each bean for classification.
13 . The bean defect sorter of claim 11 , wherein the at least one disk plate comprises:
a disk lower plate that defines a plurality of through-holes; and
a disk upper plate that is coupled to the disk lower plate, the disk upper plate comprising a plurality of partition walls that define the plurality of seating spaces in fluid communication with the plurality of through-holes, respectively.