Object/region detection and classification system with improved computer memory efficiency

1. An object detection and classification system with improved computer memory efficiency, comprising computer components and program commands executable by the computer components to implement the following units:

an input value generation unit to receive an input image and generate an input value including feature information, the input image including an object intended to be detected and classified;

a memory value generation unit to receive a reference image and generate a memory value including feature information, the reference image being associated with the object;

a memory management unit to select information having high importance from the memory values and store the selected information in a computer memory;

an aggregated value generation unit configured to:

compute similarity between the input value generated from the input image and the memory value stored in the computer memory;

calculate a weighted sum of memory values to generate an integrated value; and

aggregate the integrated value and the input value to generate an aggregated value; and

an object detection unit to detect or classify the object from the input image using the aggregated value.

2. The system according to claim 1 , wherein the aggregated value generation unit configured to:

calculate a similarity vector by computing the similarity between the input value and the memory value grid-wise; and

calculate the weighted sum of memory values grid-wise based on the similarity vector and convert the weighted sum to an integrated value.

3. The system according to claim 1 , wherein the memory management unit stores a newly-generated memory value in the computer memory only when an importance of the newly-generated memory value is equal to or higher than a predetermined value, to increase memory efficiency, and

in case of a newly-generated memory value that is stored to the computer memory when the computer memory has a limited size, the memory management unit deletes a memory value having a highest similarity or an oldest memory value among already stored memory values from the computer memory.

4. The system according to claim 3 , wherein as a similarity of the newly-generated memory value with the already stored memory values is lower and a larger amount of useful information suited for purpose is included in the image, the importance of the newly-generated memory value is measured higher.

5. The system according to claim 1 , wherein the input value generation unit includes:

an input unit encoder implemented as a machine-learning model for compressing the input image into high-level information; and

an input value model for receiving output of the input unit encoder and generating an input value including high-level feature information, and

the memory value generation unit includes:

a memory unit encoder implemented as a machine-learning model for compressing the reference image into high-level information; and

a memory value model for receiving output of the memory unit encoder and generating a memory value including high-level feature information.

6. The system according to claim 5 , wherein the input unit encoder and the memory unit encoder are configured as a same machine-learning model sharing weights.

7. The system according to claim 5 , wherein the input value model and the memory value model are configured as a same value model sharing weights.

8. A method for detecting or classifying an object in a high-resolution image, comprising:

receiving a high-resolution input image with resolution of a predetermined value or more;

reducing the input image at a predetermined ratio through multiple steps to acquire reduced images;

selecting a smallest one of the reduced images as an initial processing image;

splitting the processing image into a plurality of segments of a same size;

detecting or classifying the object in the processing image using the object detection and classification system according to claim 1 ;

determining if the processing image has a same size as the high-resolution input image;

selecting, when the processing image has a smaller size than the high-resolution input image, a segment including the object among the plurality of segments of the processing image based on a detection and classification result, and returning the selected segment;

cropping a segment corresponding to the returned segment from the reduced image of a higher level that is less reduced than the processing image;

returning the cropped segment as the processing image;

repeating a part or all of the above steps on the returned processing image; and

returning the detection and classification result to output when the processing image has a same size as the high-resolution input image.

9. A computer program recorded on a non-transitory computer-readable recording medium included in a computer system, the computer program for causing the computer system to perform the method according to claim 8 .