Knowledge distillation method for compressing image segmentation model and computing device for performing the same

1 . A knowledge distillation method for compressing an image segmentation model that is performed in a computing device including one or more processors and a memory storing one or more programs executed by the one or more processors, the knowledge distillation method comprising:

training a first image segmentation model including a first image encoder, a first image embedding layer, and a first mask decoder;

constructing a second image segmentation model including a second image encoder, a second image embedding layer, and a second mask decoder according to preset constraints; and

performing knowledge distillation for the second image embedding layer of the second image segmentation model based on the trained first image segmentation model,

wherein the performing of the knowledge distillation further includes:

outputting a first image embedding vector using the first image embedding layer based on a value output from the first image encoder, by using training data as input;

outputting a second image embedding vector using the second image embedding layer based on a value output from the second image encoder, by using the training data as input; and

training the second image embedding layer so that a distance between the first image embedding vector and the second image embedding vector is minimized.

2 . The knowledge distillation method of claim 1 , wherein the second image segmentation model is a lighter model than the first image segmentation model.

3 . The knowledge distillation method of claim 1 , wherein the first mask decoder includes a prompt encoder configured to receive at least one of a dot, a box, and text and output a prompt embedding vector,

the first image embedding layer outputs an image embedding vector based on a value output from the first image encoder by using a training image as input, and

the first mask decoder generates a mask for segmenting an image based on the output image embedding vector and the output prompt embedding vector.

4 . The knowledge distillation method of claim 1 , wherein the constructing of the second image segmentation model further includes:

constructing the second image encoder and the second image embedding layer according to the preset constraints; and

constructing the second mask decoder to be identical to the first mask decoder of the trained first image segmentation model by copying the first mask decoder.

5 . A computing device comprising:

one or more processors;

a memory; and

one or more programs,

wherein the one or more programs are configured to be stored in the memory and executed by the one or more processors, and

the one or more programs include:

instructions for training a first image segmentation model including a first image encoder, a first image embedding layer, and a first mask decoder;

instructions for constructing a second image segmentation model including a second image encoder, a second image embedding layer, and a second mask decoder according to preset constraints; and

instructions for performing knowledge distillation for the second image embedding layer of the second image segmentation model based on the trained first image segmentation model,

wherein the instructions for performing knowledge distillation include:

instructions for outputting a first image embedding vector using the first image embedding layer based on a value output from the first image encoder, by using training data as input;

instructions for outputting a second image embedding vector using the second image embedding layer based on a value output from the second image encoder, by using the training data as input; and

instructions for training the second image embedding layer so that a distance between the first image embedding vector and the second image embedding vector is minimized.

6 . The computing device of claim 5 , wherein the second image segmentation model is a lighter model than the first image segmentation model.

7 . The computing device of claim 5 , wherein the instructions for constructing a second image segmentation model include:

instructions for constructing the second image encoder and the second image embedding layer according to the preset constraints; and

instructions for constructing the second mask decoder to be identical to the first mask decoder of the trained first image segmentation model by copying the first mask decoder.