The effective description method for hand gesture is the most important in intelligent coordination assembly process based on human computer interaction. And effective hand finger knuckle detection is beneficial to the description of hand gesture.The structure characteristics of hand knuckles image are fuzzy and it is difficult to feature modeling. The extraction and learning method of excursion characteristic for hand knuckles image was presented and the hand knuckle was recognized by hand image based on Log Gaussian Cox random image model theory. The approximations of image excursion representation were given combined with level set decomposition of random image when the priori hypothesis was absented in Cox process image model. On the basis of nonparametric kernel estimation of image gray distribution, excursion characteristic was enhanced by nonlinear anisotropic filtering. And the Bayesian form of excursion measurement was established. The model learning and feature fusion algorithm on excursion characteristics with different excursion parameters was presented. And the features fusion representation of hand knuckle image was acquired. The hand knuckles image recognition results with many different hierarchical excursion data models were compared. The knuckle detection algorithm on hand image was presented. The ROC curves statisical law of hand knuckles detection with defferent models showed that the classification ablility of this method was correct and stable.The results also showed that the knuckle recognition ability of the model had some difference for different knuckle categories, and there were some differences in the deep distribution of image data between far knuckles and mid-knuckles. And the method was feasible.