The inherent “black-box” nature of neural network models leads to opaque internal logic and behavioral decision， restricting their application in critical areas. To address this issue， based on the Layer-wise Relevance Propagation （LRP） method， the interpretability of image classification models such as VGG16 and ResNet50 was researched， and an interpretation method for image classification models named SSGLRP （Selective Softmax Gradient for Layer-wise Relevance Propagation） was proposed. By introducing activation values for the positive gradients of output neurons and modifying the initial relevance values for non-target classes， the problem of LRP heatmaps containing noise and lacking class discrimination was effectively solved. The SSGLRP method was quantitatively evaluated using maximum patch masking and fixed-point game experiments， with classic LRP， Selective Layer-wise Relevance Propagation （SLRP）， and Softmax Gradient Layer-wise Relevance Propagation （SGLRP） as baseline methods. The results of the maximum patch masking experiments show that on VGG16， the prediction change obtained by SSGLRP is on average 91.8%， 60.3%， and 34.2% higher than those obtained by LRP， SLRP， and SGLRP， respectively； on ResNet50， the improvements are 61.9%， 64.9%， and 33.2%， respectively. The SSGLRP method has higher class discrimination ability and less noise with superior performance in interpreting VGG16 and ResNet50 models.