Learning data generation apparatus, learning data generation method, and recording medium
A discriminative mode generation unit generates a discriminative model that discriminates a target object included in an image using a target object image, which includes the target object and to which a label of the target object is applied, and a non-target object image, which includes a non-target object. An interest area detection unit inputs the target object image to the discriminative model, and detects an interest area which is an area that is predicted to include the target object. A rectangle generation unit generates a rectangle circumscribing the target object based on the interest area. A learning data output unit outputs learning data including a label of the target object and the rectangle.
1. A data processing apparatus comprising:
a memory storing instructions; and
one or more processors configured to execute the instructions to:
generate a discriminative model that discriminates a target object included in an image using a target object image, which includes the target object and to which a label of the target object is applied, and a non-target object image, which includes a non-target object;
input the target object image to the discriminative model, and detect an interest area which is an area that is predicted to include the target object;
generate a rectangle circumscribing the target object based on the interest area; and
output learning data including a label of the target object and the rectangle.
2. The learning data generation apparatus according to claim 1 , wherein the processor is further configured to generate a mask image by masking the rectangle in the target object image, and re-input the mask image to the discriminative model,
wherein when the mask image is recognized as the target object by the discriminative model, the processor detects the interest area from the mask image, and the processor regenerates a rectangle based on the interest area detected from the mask image.
3. The learning data generation apparatus according to claim 1 , wherein the processor is further unit configured to generate a clipped image by cutting out the rectangle in the target object image and re-input the clipped image to the discriminative model,
wherein when the clipped image is recognized as the target object by the discriminative model,
the processor re-trains the discriminative model using the clipped image,
the processor detects the interest area from the clipped image using the re-trained discriminative model, and
the processor re-generates a rectangle based on the interest area detected from the clipped image.
4. The learning data generation apparatus according to claim 1 , wherein the processor generates a feature map corresponding to the entire target object image based on a feature amount which the discriminative model generates when the target object image is input, and detects, as the interest area, an area including a value equal to or greater than a threshold value in the feature map.
5. A learning data generation method, comprising:
generating a discriminative model that discriminates a target object included in an image using a target object image which includes the target object and to which a label of the target object is applied, and a non-target object image which includes a non-target object;
inputting the target object image to the discriminative model, and detecting an interest area which is an area that is predicted to include the target object;
generating a rectangle circumscribing the target object based on the interest area; and
outputting learning data including a label of the target object and the rectangle.
6. The learning data generation method according to claim 5 , further comprising generating a mask image by masking the rectangle in the target object image, and re-inputting the mask image to the discriminative model,
wherein when the mask image is recognized as the target object by the discriminative model, the interest area is detected from the mask image, and a rectangle is regenerated based on the interest area detected from the mask image.
7. The learning data generation method according to claim 5 , further comprising generating a clipped image by cutting out the rectangle in the target object image and re-inputting the clipped image to the discriminative model,
wherein when the clipped image is recognized as the target object by the discriminative model, the discriminative model is re-trained using the clipped image, the interest area is detected from the clipped image using the re-trained discriminative model, and a rectangle is re-generated based on the interest area detected from the clipped image.
8. The learning data generation method according to claim 5 , wherein a detection of the interest area is conducted by generating a feature map corresponding to the entire target object image based on a feature amount which the discriminative model generates when the target object image is input, and detecting, as the interest area, an area including a value equal to or greater than a threshold value in the feature map.
9. A non-transitory computer-readable recording medium storing a program, the program causing a computer to perform a process comprising:
generating a discriminative model that discriminates a target object included in an image using a target object image which includes the target object and to which a label of the target object is applied, and a non-target object image which includes a non-target object;
inputting the target object image to the discriminative model, and detecting an interest area which is an area that is predicted to include the target object;
generating a rectangle circumscribing the target object based on the interest area; and
outputting learning data including a label of the target object and the rectangle.