Interpretable Deep Neural Networks (DNNs) are becoming increasingly crucial, particularly in high-risk domains such as medical diagnosis and autonomous driving. Understanding the reasoning behind AI decisions is paramount. To address this need, a novel parallel interpretation network based on Semantic Visual Probe (SVProbe) and Counterfactual Verification is proposed, aiming to enhance the interpretability of deep neural networks and facilitate user comprehension of network decisions. The novelty of the network lies in that SVProbe incorporates prior knowledge into gradient computation and explains the textual attributes and their corresponding visual mappings that guide the model decisions. Concurrently, Counterfactual Verification generates counterfactual samples by altering the color, texture and shape information in the images, feeding them into the original model to evaluate the consistency of different decision interpretations, thus providing reliability verification for SVProbe’s interpretations. This parallel interpretation network structure enables users to understand network decisions more intuitively and efficiently. Extensive experiments on pre-trained models and public datasets demonstrate that our method effectively reveals the evidence behind model decisions and outperforms existing methods in terms of interpretability and credibility. Specifically, on VGG16, ResNet50 and ResNet101, the Bounding box (Bbox) scores are 64.85%, 77.02% and 77.26%, respectively. Compared with other methods, the method improves the attribute accuracy by an average of 1.8% and the top 1 attribute accuracy by 1.9%.