Automated close-range photogrammetric network orientation has been traditionally performed with the use of coded targets in the object space to allow for initial point correspondence determination and subsequent network orientation. Feature-based matching (FBM) techniques have recently offered an alternative procedure for point correspondence calculation between image pairs. FBM algorithms, however, do not come free of complications. Due to the way that FBM considers point correspondences based on the similarity of feature descriptors, a considerable number of mismatches (outliers) can be anticipated, especially with increasing angles of convergence between images. For the critical component of initial Relative Orientation, it is essential that outliers are detected and largely removed from the matched point data. This paper reports on the application of a machine-based learning approach to outlier detection in FBM. The method of Support Vector Regression is evaluated and compared to other outlier removal algorithms for cases of convergent image configurations. Various experimental tests were conducted in controlled networks and with other real datasets using the ‘Identifying point correspondences by Correspondence Function’ (ICF) algorithm, employing different SVR kernel functions. The paper also reports on optimisations made to achieve better results when highly convergent imaging geometries are adopted.