The Impact of a Number of Samples on Unsupervised Feature Extraction, Based on Deep Learning for Detection Defects in Printed Circuit Boards

Deep learning provides new ways for defect detection in automatic optical inspections (AOI). However, the existing deep learning methods require thousands of images of defects to be used for training the algorithms. It limits the usability of these approaches in manufacturing, due to lack of images of defects before the actual manufacturing starts. In contrast, we propose to train a defect detection unsupervised deep learning model, using a much smaller number of images without defects. We propose an unsupervised deep learning model, based on transfer learning, that extracts typical semantic patterns from defect-free samples (one-class training). The model is built upon a pre-trained VGG16 model. It is further trained on custom datasets with different sizes of possible defects (printed circuit boards and soldered joints) using only small number of normal samples. We have found that the defect detection can be performed very well on a smooth background; however, in cases where the defect manifests as a change of texture, the detection can be less accurate. The proposed study uses deep learning self-supervised approach to identify if the sample under analysis contains any deviations (with types not defined in advance) from normal design. The method would improve the robustness of the AOI process to detect defects.

Interpretation of clearness day-ahead forecast errors using novel cloud classification

<p>        The formation and dissipation of clouds are one of the longest studied and yet least understood phenomenon in nature. This is crucial in atmospheric and climate science as clouds have a significant impact on radiative forcing. In numerical weather prediction, solar radiation forecasts have lower skill than other parameters as temperature forecasts despite recent progresses. This study aims at better understanding cloud situations over Europe and how solar radiation forecast errors are related to these situations. Therefore, an enhanced cloud class algorithm based on unsupervised Deep Learning and hierarchical clustering is introduced. By using the MODIS optical cloud thickness product, the algorithm is able to classify 14 different daily cloud situations which are applied on defined tile regions (approximately 70,000 km²) of Europe. These different classes differ in both optical cloud phase and the overall structure of the cloud shape. The usefulness of the cloud classes is illustrated by showing regional differences of cloud type frequencies over the last 20 years. To better understand solar radiation forecast errors, the cloud classes are assigned to ECMWF IFS clearness day-ahead forecast errors. We show that high-water content and mixed-cloud phase situations lead to highest absolute forecast errors for single sites. Summed up over an area, we observe an accumulation of forecast errors for mixed-cloud phase situations whereas for other cloud situations forecast errors are more likely to cancel each other out (e.g. broken high-water content clouds). This study is useful for researchers and practitioners to better understand situations of high solar radiation errors by using the developed cloud product.</p>

De-noising Receiver function data using the unsupervised deep learning approach

Summary The converted wave data (P-to-s or S-to-p), traditionally termed as receiver functions, are often contaminated with noise of different origin that may lead to the erroneous identification of phases and thus influence the interpretations. Here we utilize an unsupervised deep learning approach called Patchunet to de-noise the converted wave data. We divide the input data into several patches, which are input to the encoder and decoder network to extract some meaningful features. The method de-noises an image patch-by-patch and utilizes the redundant information on similar patches to obtain the final de-noised results. The method is first tested on a suite of synthetic data contaminated with various amount of Gaussian and realistic noise and then on the observed data from three permanent seismic stations: HYB (Hyderabad, India), LBTB (Lobatse, Botswana, South Africa), COR (Corvallis, Oregon, USA). The method works very well even when the signal-to-noise ratio is poor or with the presence of spike noise and deconvolution artifacts. The field data demonstrate the effectiveness of the method for attenuating the random noise especially for the mantle phases, which show significant improvements over conventional receiver function based images.

Unsupervised deep learning based ego motion estimation with a downward facing camera

Knowing the robot's pose is a crucial prerequisite for mobile robot tasks such as collision avoidance or autonomous navigation. Using powerful predictive models to estimate transformations for visual odometry via downward facing cameras is an understudied area of research. This work proposes a novel approach based on deep learning for estimating ego motion with a downward looking camera. The network can be trained completely unsupervised and is not restricted to a specific motion model. We propose two neural network architectures based on the Early Fusion and Slow Fusion design principle: “EarlyBird” and “SlowBird”. Both networks share a Spatial Transformer layer for image warping and are trained with a modified structural similarity index (SSIM) loss function. Experiments carried out in simulation and for a real world differential drive robot show similar and partially better results of our proposed deep learning based approaches compared to a state-of-the-art method based on fast Fourier transformation.