Quality inspection is inevitable in the steel industry so there are already benchmark datasets for the visual inspection of steel surface defects. In our work, we show, contrary to previous recent articles, that a generic state-of-art deep neural network is capable of almost-perfect classification of defects of two popular benchmark datasets. However, in real-life applications new types of errors can always appear, thus incremental learning, based on very few example shots, is challenging. In our article, we address the problems of the low number of available shots of new classes, the catastrophic forgetting of known information when tuning for new artifacts, and the long training time required for re-training or fine-tuning existing models. In the proposed new architecture we combine EfficientNet deep neural networks with randomized classifiers to aim for an efficient solution for these demanding problems. The classification outperforms all other known approaches, with an accuracy 100% or almost 100%, on the two datasets with the off-the-shelf network. The proposed few-shot learning approach shows considerably higher accuracy at a low number of shots than the different methods under testing, while its speed is significantly (at least 10 times) higher than its competitors. According to these results, the classification and few-shot learning of steel surface defects can be solved more efficiently than was possible before.

No abstract

Our paper deals with active multiview object recognition focusing on the directional support of sequential multiple shots. Since inertial sensors are easily available nowadays, we propose the use of them to estimate the orientation change of the camera and thus to estimate the probability of relative poses. With the help of relative orientation change, we can compute transition probabilities between possible poses and can use a hidden Markov model to evaluate state (pose) sequences and can thus increase the recognition rate. Furthermore, we can plan our next viewing position to minimize the risk of misclassification, resulting in higher overall recognition rates. Besides giving the theoretical details, we use two datasets to illustrate the performance of our model through several tests including occlusion, blur, Gaussian noise, and to compare to a solution with a long short-term memory network.