I have created a process that uses a neural network (NN) to evaluate images. The images are from network cameras and I want the NN to classify the images into two groups: images in which people and/or vehicles appear or images in which people and/or vehicles do not appear. (The process then emails me whenever it finds an image in which people or vehicles appear.) I'm using the ImageMiner-1.4.1 extension. I am getting poor performance from my model and I assume it is because I am not really extracting the best features from the images so that the NN can make accurate classifications. I would appreciate it if anyone who is familiar with the ImageMiner-1.4.1 extension could give me some suggestions as to the types of features I should be using for this kind of task (and any other suggestions that might be helpful).
(In case anyone is wondering what I mean by poor performance. I've used cross-validation and my best performance is a precision of 86% [not too bad] and accuracy of 27% [ugh!!]. The model performs much worse than this when applied against other unlabeled images. I also tried using the "similarity" operator (provided by the ImageMiner-1.4.1 extension and it gave me FABULOUS performance on cross-validation (100% precision and 100% accuracy) - but then when I actually used it against more unlabeled images, it failed miserably. It only identified about 30% correctly.)
Thanks in advance.