I have a lot of experience on what you want, some of image processing and computer vision . I recetly finished my master of science (UADY, PNPC degrees with high quality in México), in this degree i worked on computer vision and machine learning, where i developed a system capable to estimate the head pose and gaze of the peple in a long distance, i also used in my work some neural networks algorithms. I also in the Masters i took many cv subjects like: computer vision, 3D vision, image processing, machine learning so i think i have good support and background for this project
At first glance i think that i can develop a very precise and robust system to detect that crosses with a machine learning algorithm like SVM (since is binary classification and this is a very powerful and fast method) and to train the algorithm we can use some very useful descriptor like HOG (histogram oriented gradient) so first of all we train the support vector machine with several images of different kind of crosses (large, small, different thikness, rotated) and extract the useful information with the HOG descrptors, and the rest is just make some test to tune some parameters of the SVM or HOG.