Research on Quantitative Evaluation Method of Interpretability Based on Deep Learning Model

Since deep learning models are often regarded as “black box models” and their decision-making process lacks transparency, this paper introduces a new quantitative evaluation framework for the interpretability of deep learning models, aiming to improve the interpretability of models. The framework is based on YOLOv8 and MobileSAM, and introduces two quantitative evaluation indicators: $\mathrm{N}-\mathrm{IOU}$ and N -Recall. The $\mathrm{N}-\mathrm{IOU}$ index extends the traditional concept of IOU by measuring the accuracy and comprehensiveness of the model explanation; The $\mathbf{N}$-Recall metric focuses on the recall of the model when identifying positive samples, especially the performance of difficult samples. Experiments show that the dataset selected using these indicators is similar to the original dataset in terms of model loss and classification accuracy, while the randomly selected dataset has an accuracy difference of $1.4-1.6 \%$. In addition, this study also provides a method for automating the generation of structured annotated data, which reduces the cost and time of data preparation, and is of great significance for the construction and maintenance of large-scale image datasets.