In order to achieve nondestructive detection of heavy metal cadmium in lettuce leaves, computer vision technology was used as the research method, which combined image processing method and feature selection method, to identify four gradients of heavy metal cadmium stress lettuce leaves. First of all, the leaf image of lettuce was obtained by digital camera. Then, the Kmeans clustering algorithm was used to segment the image, and the color, shape and texture of the image were extracted from the extracted target image. A total of 46 image features were obtained. In order to make the model easier and reduce the amount of data, the image feature was dimensioned by competitive adaptive reweighted sampling (CARS) and variable importance analysis based on random variable combination (VIAVC). The partial least squares discriminant analysis (PLS-DA) and random forest (RF) were used to construct the model for identification of cadmium stress in lettuce. The results showed that in the seven combined feature models, the optimal model was given by the model of color, shape and texture fusion. The accuracy of the training set classification was 92%. The color, shape and texture fusion features were reduced by CARS and VIAVC, and it was found that the dimensionality and visualization of VIAVC were better than those of CARS. Using the reduced dimension of the lowdimensional mapping point to build the model, the accuracy of the training set classification and accuracy of the prediction set of RF model were higher than those of the PLS-DA. Among them, the accuracy of the training set and predictive set classification based on VIAVC dimensionality reduction were 98.0% and 96.0%, respectively. It can be seen that the RF model based on VIAVC dimensionality can better identify the lettuce leaves with different cadmium stress levels under the premise of greatly reducing the feature dimension.