Training device and training method
A training method, comprising: acquiring pieces of training image information each indicating a training image; generating one-dimensional training image signals respectively corresponding to the pieces of training image information by making the training image indicated by each of the pieces of training image information into one dimension; generating a matrix in which the one-dimensional training image signals are arranged in parallel with each other; calculating a right singular vector and a singular value by performing singular value decomposition on the matrix; generating a trained model that is based on the right singular vector and the singular value and outputs an inference result using, as an explanatory variable, a one-dimensional inference target image signal indicating an image obtained by making an inference target image that is an image obtained by image-capturing an inference target object into one dimension; and outputting the trained model as trained model information.
1 . A training device, comprising processing circuitry;
to acquire a plurality of pieces of training image information each indicating a training image;
to generate a plurality of one-dimensional training image signals respectively corresponding to the plurality of pieces of training image information by making the training image indicated by each of the plurality of pieces of training image information into one dimension;
to generate a matrix in which the plurality of one-dimensional training image signals is arranged in parallel with each other;
to calculate a right singular vector and a singular value by performing singular value decomposition on the matrix;
to generate a trained model that is based on the right singular vector and the singular value and outputs an inference result using, as an explanatory variable, a one-dimensional inference target image signal indicating an image obtained by making an inference target image that is an image obtained by image-capturing an inference target object into one dimension; and
to output the trained model as trained model information.
2 . The training device according to claim 1 , wherein
the processing circuitry generates the trained model that outputs, as the inference result, a one-dimensional approximate signal that is an approximate signal of the one-dimensional inference target image signal on a basis of the right singular vector and the singular value when the one-dimensional inference target image signal is input as the explanatory variable.
3 . The training device according to claim 1 , wherein
when the one-dimensional inference target image signal is input as the explanatory variable, the processing circuitry generates the trained model that outputs a feature vector of the inference target image corresponding to the one-dimensional inference target image signal as the inference result on a basis of the right singular vector and the singular value.
4 . The training device according to claim 1 , wherein
in a case where a left singular vector which is calculated by performing singular value decomposition on the matrix, the right singular vector, and the singular value are already obtained, as a method of singular value decomposition of a matrix obtained by connecting, in a row direction, the matrix which is subjected to singular value decomposition to obtain the left singular vector, the right singular vector, and the singular value and the matrix, the processing circuitry calculates the right singular vector and the singular value by performing a matrix operation of updating the left singular vector that is already obtained, the right singular vector that is already obtained, and the singular value that is already obtained using the matrix.
5 . The training device according to claim 1 , wherein
in a case where a left singular vector which is calculated by performing singular value decomposition on the matrix, the right singular vector, and the singular value are already obtained, the processing circuitry selects one one-dimensional training image signal from the matrix which is subjected to the singular value decomposition to obtain the left singular vector, the right singular vector, and the singular value, and generates a matrix excluding the one-dimensional training image signal selected from the matrix which is subjected to the singular value decomposition; and
as a method of obtaining a singular value decomposition result of the matrix excluding the one-dimensional training image signal selected, the processing circuitry calculates the right singular vector and the singular value by performing a matrix operation of updating the left singular vector that is already obtained, the right singular vector that is already obtained, and the singular value that is already obtained using the one-dimensional training image signal selected.
6 . The training device according to claim 1 , wherein
the processing circuitry generates the one-dimensional training image signals by raster scanning the training image indicated by the training image information.
7 . The training device according to claim 1 , wherein
the processing circuitry acquires the plurality of pieces of training image information respectively indicating a plurality of training images having equal numbers of pixels.
8 . The training device according to claim 1 , wherein
the processing circuitry acquires, as the training image information, a plurality of pieces of partial image information respectively indicating a plurality of partial images obtained by dividing a captured image in which a target object is captured.
9 . The training device according to claim 8 , wherein
the processing circuitry generates the matrix by arranging, in parallel with each other, the plurality of one-dimensional training image signals respectively corresponding to pieces of partial image information, out of the plurality of pieces of partial image information, the pieces of partial image information respectively satisfying predetermined conditions.
10 . The training device according to claim 9 , wherein
the processing circuitry generates the matrix by arranging, in parallel with each other, the one-dimensional training image signals on a basis of the plurality of pieces of partial image information corresponding to a same image region in the captured image.
11 . The training device according to claim 9 , wherein the processing circuitry further performs:
clustering of the plurality of partial images respectively indicated by the plurality of pieces of partial image information; and
generating the matrix by arranging the plurality of one-dimensional training image signals in parallel with each other for each of the plurality of partial images belonging to each class classified on a basis of a result of the clustering.
12 . The training device according to claim 9 , wherein the processing circuitry
generates, on a basis of the predetermined conditions, the matrix corresponding to each of the predetermined conditions,
calculates the right singular vector and the singular value for the matrix corresponding to each of the conditions, and
generates the trained model corresponding to the matrix which corresponds to each of the conditions on a basis of the right singular vector and the singular value corresponding to the matrix which corresponds to each of the conditions.
13 . A training method, comprising:
acquiring a plurality of pieces of training image information each indicating a training image;
generating a plurality of one-dimensional training image signals respectively corresponding to the plurality of pieces of training image information by making the training image indicated by each of the plurality of pieces of training image information into one dimension;
generating a matrix in which the plurality of one-dimensional training image signals is arranged in parallel with each other;
calculating a right singular vector and a singular value by performing singular value decomposition on the matrix;
generating a trained model that is based on the right singular vector and the singular value and outputs an inference result using, as an explanatory variable, a one-dimensional inference target image signal indicating an image obtained by making an inference target image that is an image obtained by image-capturing an inference target object into one dimension; and
outputting the trained model as trained model information.