TY - GEN

T1 - Visual Word based Neural Tree for Interpretable Recognition of Images

AU - Kozinets, Roman

AU - Berikov, Vladimir

N1 - The work was partly supported by RFBR grant 19-29-01175, and by the State Contract of the Sobolev Institute of Mathematics, Project No. FWNF-2022-0015.

PY - 2022

Y1 - 2022

N2 - It is necessary to have an explanation of the model prediction to increase confidence in the decision making in a broad range of image recognition tasks, especially medical image recognition.In this paper, we proposed a new method of interpretable image recognition. The basic idea was to use a combination of convolutional neural network, similarity based decision tree, and a trainable bag of visual words. During recognition, the model compared parts of the image with trained templates and made a decision based on the similarity between them. The recognition process was presented as a sequence of logical decision rules which can be easily understood by specialists in the applied area. This technique is similar to the human way of visual recognition. Testing on two datasets showed a competitive recognition quality of the proposed method, compared to a classical convolutional network ResNet50. At the same time, the method provided a possibility of interpreting the prediction.

AB - It is necessary to have an explanation of the model prediction to increase confidence in the decision making in a broad range of image recognition tasks, especially medical image recognition.In this paper, we proposed a new method of interpretable image recognition. The basic idea was to use a combination of convolutional neural network, similarity based decision tree, and a trainable bag of visual words. During recognition, the model compared parts of the image with trained templates and made a decision based on the similarity between them. The recognition process was presented as a sequence of logical decision rules which can be easily understood by specialists in the applied area. This technique is similar to the human way of visual recognition. Testing on two datasets showed a competitive recognition quality of the proposed method, compared to a classical convolutional network ResNet50. At the same time, the method provided a possibility of interpreting the prediction.

UR - https://www.scopus.com/inward/record.url?eid=2-s2.0-85147505671&partnerID=40&md5=648b886a00d430223e73d54f41c3ad3c

UR - https://www.mendeley.com/catalogue/ebb3fbf2-ad5c-3ac3-b8e7-afd80239e4cb/

U2 - 10.1109/sibircon56155.2022.10017004

DO - 10.1109/sibircon56155.2022.10017004

M3 - Conference contribution

SN - 9781665464802

SP - 1830

EP - 1835

BT - 2022 IEEE International Multi-Conference on Engineering, Computer and Information Sciences (SIBIRCON)

PB - Institute of Electrical and Electronics Engineers (IEEE)

T2 - 2022 IEEE International Multi-Conference on Engineering, Computer and Information Sciences, SIBIRCON 2022

Y2 - 11 November 2022 through 13 November 2022

ER -