Robert Bosch GmbH utilizes Automatic Optical Inspection to detect pieces that are produced with defects. Convolutional Neural Networks are state of the art classifiers. However, in order to obtain good performance the networks need to be trained on large amount of labeled data. In most of the tasks, only small size of labeled datasets are available. This work evaluates two approaches for addressing the problem of small data. The first one consist of a new kind of neural networks called Receptive Fields Networks, these models combine the idea of multi-scale image analysis and applied into CNNs. The main idea is learn complex filters from a set of basis Gaussian Derivative filers, in order to capture local variation in the images. In the second approach a decoder is added into a CNN, which will add a second objective into the overall loss, in this approach, also unlabeled data is employed to improve performance. Evaluation of this two approaches is made by measuring accuracy in the test set, false positive and false negative rates.