Electronic apparatus for training classification network and operating method thereof, and electronic apparatus using classification network

An electronic apparatus comprising:

• a memory storing a classification network that includes a plurality of feature extraction layers and is trained to classify an object included in an image;
• and
• a processor configured to:
• control the classification network to calculate a class score indicating a probability which the object belongs to each of a plurality of classes in response to receiving a input image including the object;
• calculate an activation map loss value representing a same position difference between a binary image corresponding to the input image and each of a plurality of activation maps, in response to receiving the binary image, each of the plurality of activation maps is obtained by emphasizing a unique feature in data input to each of the plurality of feature extraction layers;
• calculate a softmax loss value representing an error of the class score;
• calculate a final loss value based on the activation map loss value and the softmax loss value;
• perform a training operation for the classification network based on the final loss value, by inputting the final loss value to an output terminal of the classification network and updating at least one of a plurality of weighted parameters included in a convolution layer which is included in each of the plurality of feature extraction layers;
• and
• repeat the training operation until the final loss value becomes a threshold value or less,
• wherein the trained classification network is a neural network trained based on the final loss value which is a result of a weight calculation of the softmax loss value corresponding to the input image input to the classification network and the activation map loss value.

A method of operating an electronic apparatus, the method comprising:

• inputting a training image including an object to a classification network including a plurality of feature extraction layers;
• calculating a class score indicating a probability which the object belongs to each of a plurality of classes, which is output from the classification network;
• calculating an activation map loss value representing a same position difference between a binary image corresponding to the training image and each of a plurality of activation maps, in response to receiving the binary image, each of the plurality of activation maps is obtained by emphasizing a unique feature in data input to each of the plurality of feature extraction layers;
• calculating a softmax loss value representing an error of the class score;
• calculating a final loss value based on the activation map loss value and the softmax loss value;
• performing a training operation for the classification network based on the final loss value, by inputting the final loss value to an output terminal of the classification network and updating at least one of a plurality of weighted parameters included in a convolution layer which is included in each of the plurality of feature extraction layers;
• and
• repeating the training operation until the final loss value becomes a threshold value or less.