Image processing system and image processing method

1. An image processing system comprising:

a memory storing a training image set and a reference image set; and

a processor including hardware, wherein the processor is configured to:

performing a determination process comprising:

generate an augmented image set of endoscope images by applying data augmentation, including one or more of an image resizing process, a brightness correction process and a smoothing process, to image data of endoscope images included in the training image set, the data augmentation being applied based on a candidate augmentation parameter on the one or more of the image resizing process, the brightness correction process and the smoothing process, that is a candidate for an augmentation parameter;

calculate an augmentation feature statistic based on the augmented image set, the augmentation feature statistic being a statistic of a feature of one or more pixels in the endoscope images of the augmented image set representing a recognition target in the endoscope images of the augmented image set;

calculate a reference feature statistic based on the reference image set, the reference feature statistic being a statistic of the feature of one or more pixels in endoscope images of the reference image set representing the recognition target in the endoscope images of the reference image set; and

determine whether the augmentation feature statistic and the reference feature statistic have a predefined similarity;

repeat the determination process with an incremented candidate augmentation parameter as the candidate for the augmentation parameter until the augmentation feature statistic of one of the augmented image sets generated based on the candidate augmentation parameter and the incremented candidate augmentation parameters and the reference feature statistic has the predefined similarity; and

generate a trained model for image recognition process based on the endoscope images of the augmented image set having the augmentation parameter that has the predefined similarity with the reference feature statistic.

2. The image processing system as defined in claim 1 ,

wherein the reference image set includes at least one endoscope image among the endoscope images corresponding to frames including a frame in which the recognition target has appeared and frames within a predetermined time period after the frame.

3. The image processing system as defined in claim 2 ,

wherein the endoscope images of the reference image set include images obtained by capturing an interior of a tubular structure.

4. The image processing system as defined in claim 1 ,

wherein the processor is configured to repeat the determination process a predetermined number of times, and

wherein the predefined similarity is determined based on one of the candidate augmentation parameter and the one or more incremented candidate augmentation parameters used in the repeated determination process has a corresponding augmentation feature statistic with a highest similarity to the reference feature statistic.

5. The image processing system as defined in claim 1 ,

wherein the recognition target in an endoscope image included in the endoscope images of the reference image set is an object of interest positioned in a peripheral area in the endoscope image.

6. The image processing system as defined in claim 1 ,

wherein the data augmentation includes the image resizing process, and

wherein the augmentation parameter that has the predefined similarity with the reference feature statistic is a lower limit of an image resizing rate based on the similarity between the augmentation feature statistic and the reference feature statistic, the lower limit of the image resizing rate representing a limit to image size reduction in the image resizing process.

7. The image processing system as defined in claim 1 ,

wherein the data augmentation includes the brightness correction process, and

wherein the augmentation parameter that has the predefined similarity with the reference feature statistic is a lower limit of brightness correction based on the similarity between the augmentation feature statistic and the reference feature statistic, the lower limit of the brightness correction representing a limit to brightness reduction in the brightness correction process.

8. The image processing system as defined in claim 1 ,

wherein the data augmentation includes the smoothing process, and

wherein the augmentation parameter that has the predefined similarity with the reference feature statistic is an upper limit of intensity of the smoothing process based on the similarity between the augmentation feature statistic and the reference feature statistic.

9. The image processing system as defined in claim 1 ,

wherein the processor is configured to use at least one of a size, brightness, and a blur amount of the recognition target, as the feature of the recognition target.

10. The image processing system as defined in claim 1 ,

wherein the processor is configured to use at least one of a frequency distribution, a mean value, a variance, a maximum value, a minimum value, and a higher-order moment of the feature of the recognition target, as the statistic of the feature of the recognition target.

11. The image processing system as defined in claim 1 ,

wherein the processor is configured to generate the trained model through machine learning based on the endoscope images of the augmented image set having the augmentation parameter that has the predefined similarity with the reference feature statistic.

12. An image detection system comprising:

a memory storing the trained model generated by the image processing system according to claim 1 ; and

a processor including hardware, wherein the processor is configured to detect the recognition target in an acquired applied endoscope image by performing processing thereon based on the trained model.

13. The image detection system as defined in claim 12 ,

wherein the applied image is an in-vivo endoscope image capturing an interior of a living body, and

wherein the recognition target is an object of interest captured in the in-vivo endoscope image.

14. An image processing method comprising:

acquiring a training image set and a reference image set;

performing a determination process comprising:

generating an augmented image set of endoscope images by applying data augmentation, including one or more of an image resizing process, a brightness correction process and a smoothing process, to image data of endoscope images included in the training image set, the data augmentation being applied based on a candidate augmentation parameter on the one or more of the image resizing process, the brightness correction process and the smoothing process, that is a candidate for an augmentation parameter;

calculating an augmentation feature statistic based on the augmented image set, the augmentation feature statistic being a statistic of a feature of one or more pixels in the endoscope images of the augmented image set representing a recognition target in the endoscope images of the augmented image set;

calculating a reference feature statistic based on the reference image set, the reference feature statistic being a statistic of the feature of one or more pixels in endoscope images of the reference image set representing the recognition target in the endoscope images of the reference image set; and

determining whether the augmentation feature statistic and the reference feature statistic has a predefined similarity;

repeating the determination process with an incremented candidate augmentation parameter as the candidate for the augmentation parameter until the augmentation feature statistic of one of the augmented image sets generated based on the candidate augmentation parameter and the incremented candidate augmentation parameters and the reference feature statistic has the predefined similarity; and

generating a trained model for image recognition process based on the endoscope image of the augmented image set having the augmentation parameter that has the predefined similarity with the reference feature statistic.

15. An image detection system comprising:

a processor including hardware, wherein the processor is configured to:

receive an acquired endoscope image; and

detect the recognition target in the acquired endoscope image based on the trained model generated by the image processing system according to claim 1 .