|
An Explainability Guided Method for Enhancing Deep Neural Networks' Focus on the Primary Object in Image Classification
|
Ali Jafari *   , Sharif Hasani |
|
|
|
Abstract: (2 Views) |
With the advancement of the use of deep learning in sensitive and critical industries such as military command and control systems, the existence of elements such as reliability, the ability to interpret results, and understanding the decision-making structure of trained models has become more and more felt. For this purpose, solutions such as model explainability algorithms such as Grad-CAM, SHAP, LIME, etc. have been proposed. In recent years, an approach to adding explainability guidance during the model training process, Explanation-Guided Learning (EGL), has also been proposed. The aim is to increase the prediction accuracy and gain user confidence in the model results. In this research, a new solution using the EGL approach is presented to increase the focus of classification models on the main subject. The proposed objective function has terms for classification, focusing on the main subject, and reducing attention to unimportant points in the image. The proposed method has been retrained on the ResNet18 network with the TinyImageNet dataset, and its results have been improved in two aspects: prediction accuracy and explainability quality. Also, compared to previous research, the proposed method has increased the network's focus on the main subject and reduced the model's focus on background features.
|
|
| Keywords: Deep Neural Networks, Explainability, Interpretability, Explanation-Guided Learning EGL, Image Classification |
|
|
Full-Text [PDF 1551 kb]
(3 Downloads)
|
Type of Study: Research |
Subject:
Artificial Intelligence
Received: 2026/06/1 | Accepted: 2026/09/14 | Published: 2026/10/4
|
|
|
|
|
|
|
| Send email to the article author |
|
|
|
| Add your comments about this article |
|
|
|
