Deep Neural Networks (DNNs) have demonstrated impressive performance in complex machine learning tasks such as image classification and speech recognition. However, due to their multi-layer nonlinear structure, DNNs are not transparent, making it difficult to understand the reasoning behind their classification decisions for new, unseen data. Recently, several methods have been proposed to visualize the impact of individual pixels on a DNN's classification decision, creating heatmaps that show pixel importance. While heatmaps can provide intuitive insights, there is a lack of objective quality measures to evaluate their effectiveness. This paper presents a general methodology based on region perturbation to evaluate ordered collections of pixels, such as heatmaps. The authors compare heatmaps computed by three different methods—sensitivity analysis, deconvolution, and Layer-wise Relevance Propagation (LRP)—on three large datasets: SUN397, ILSVRC2012, and MIT Places. The main finding is that LRP provides better explanations of DNN classification decisions compared to the other two methods. The paper also discusses the theoretical basis for this result and explores the practical implications of using heatmaps for unsupervised assessment of neural network performance. The authors introduce a generic framework for evaluating heatmaps, extending the approach from binary inputs to color images. They demonstrate the effectiveness of LRP through experimental results, showing that it better identifies relevant pixels and provides less noisy heatmaps. The paper also investigates the correlation between heatmap quality and neural network performance, suggesting that heatmaps may be useful for assessing network performance. In conclusion, the paper contributes to the understanding and transparency of DNNs by providing a method to quantify the quality of heatmaps, which can be used for better intuition about what the network has learned and for prioritizing image regions for further analysis.Deep Neural Networks (DNNs) have demonstrated impressive performance in complex machine learning tasks such as image classification and speech recognition. However, due to their multi-layer nonlinear structure, DNNs are not transparent, making it difficult to understand the reasoning behind their classification decisions for new, unseen data. Recently, several methods have been proposed to visualize the impact of individual pixels on a DNN's classification decision, creating heatmaps that show pixel importance. While heatmaps can provide intuitive insights, there is a lack of objective quality measures to evaluate their effectiveness. This paper presents a general methodology based on region perturbation to evaluate ordered collections of pixels, such as heatmaps. The authors compare heatmaps computed by three different methods—sensitivity analysis, deconvolution, and Layer-wise Relevance Propagation (LRP)—on three large datasets: SUN397, ILSVRC2012, and MIT Places. The main finding is that LRP provides better explanations of DNN classification decisions compared to the other two methods. The paper also discusses the theoretical basis for this result and explores the practical implications of using heatmaps for unsupervised assessment of neural network performance. The authors introduce a generic framework for evaluating heatmaps, extending the approach from binary inputs to color images. They demonstrate the effectiveness of LRP through experimental results, showing that it better identifies relevant pixels and provides less noisy heatmaps. The paper also investigates the correlation between heatmap quality and neural network performance, suggesting that heatmaps may be useful for assessing network performance. In conclusion, the paper contributes to the understanding and transparency of DNNs by providing a method to quantify the quality of heatmaps, which can be used for better intuition about what the network has learned and for prioritizing image regions for further analysis.