Search

Fusion of spectral and spatial information has been proved to be an effective approach to improve model performance in near-infrared hyperspectral data analysis. Regardless, most of the existing spectral-spatial classification methods require fairly complex pipelines and exact selection of parameters, which mainly depend on the investigator's experience and the object under test. Convolutional neural network (CNN) is a powerful tool for representing complicated data and usually works with few "hand-engineering", making it an appropriate candidate for developing a general and automatic approach. In this paper, a two-branch convolutional neural network (2B–CNN) was developed for spectral-spatial classification and effective wavelengths (EWs) selection. The proposed network was evaluated by three classification data sets, including herbal medicine, coffee bean and strawberry. The results showed that the 2B–CNN obtained the best classification accuracies (96.72% in average) when compared with support vector machine (92.60% in average), one dimensional CNN (92.58% in average), and grey level co-occurrence matrix based support vector machine (93.83% in average). Furthermore, the learned weights of the two-dimensional branch in 2B–CNN were adopted as the indicator of EWs and compared with the successive projections algorithm. The 2B–CNN models built with wavelengths selected by the weight indicator achieved the best accuracies (96.02% in average) among all the examined EWs models. Different from the conventional EWs selection method, the proposed algorithm works without any additional retraining and has the ability to comprehensively consider the discriminative power in spectral domain and spatial domain. Image 1 • Convolutional neural network is developed for hyperspectral data classification. • Two-branch network structure is adopted for spectral-spatial information fusion. • The learned weights are adopted as the indicator of effective wavelengths. • The proposed network shows robustness under small sample size. [ABSTRACT FROM AUTHOR]

The ensemble pruning system is an effective machine learning framework that combines several learners as experts to classify a test set. Generally, ensemble pruning systems aim to define a region of competence based on the validation set to select the most competent ensembles from the ensemble pool with respect to the test set. However, the size of the ensemble pool is usually fixed, and the performance of an ensemble pool heavily depends on the definition of the region of competence. In this paper, a dynamic pruning framework called margin-based Pareto ensemble pruning is proposed for ensemble pruning systems. The framework explores the optimized ensemble pool size during the overproduction stage and finetunes the experts during the pruning stage. The Pareto optimization algorithm is used to explore the size of the overproduction ensemble pool that can result in better performance. Considering the information entropy of the learners in the indecision region, the marginal criterion for each learner in the ensemble pool is calculated using margin criterion pruning, which prunes the experts with respect to the test set. The effectiveness of the proposed method for classification tasks is assessed using datasets. The results show that margin-based Pareto ensemble pruning can achieve smaller ensemble sizes and better classification performance in most datasets when compared with state-of-the-art models. [ABSTRACT FROM AUTHOR]

[Display omitted] • Detecting strawberry anomalies using hyperspectral imaging. • Constructing models by joint optimization of autoencoder and self-supervised learning. • Simulating spectra with various objects to build effective self-supervised classifier. • Visualizing strawberry anomalies including bruise, infection, and contamination. • The proposed method shows good robustness against impure data in the training set Developing unsupervised anomaly detection methods for hyperspectral data is of great importance for its applications in quality and safety control. As a frequently-used anomaly detection method, the autoencoder might suffer from the ineffectiveness of extracting essential representations for distinguishing normal and anomalous samples, since it is only trained to minimize the reconstruction error. To improve the performance of the autoencoder, an anomaly detection method for hyperspectral data named SSC-AE is proposed based on the joint learning of autoencoder and self-supervised classifier, and it is evaluated on the detection of quality defects of strawberries, including bruise, fungal infection, and soil contamination. In the proposed architecture, a self-supervised classification task was designed to discriminate the low-dimensional representations of the normal data and the synthetic anomalous data that extracted from the autoencoder, consequently inducing the autoencoder to learn low-dimensional representations with more discriminative power. Experimental results on hyperspectral data of strawberries show that the SSC-AE demonstrated the best anomaly detection performance and its AUC gains compared with the one-dimensional autoencoder, one-dimensional variational autoencoder, two-dimensional autoencoder, one-class support vector machine and self-supervised classifier achieved 29.0%, 21.2%, 55.5%, 28.1% and 24.9%, respectively. It was also found that the locations and shapes of all three types of strawberry anomalies could be successfully visualized by predicting spectra pixel-by-pixel. Furthermore, the algorithm robustness against impure data in the training procedure was investigated by randomly mixing some anomalous samples into the training set. The SSC-AE degraded gracefully and outperformed all the comparison methods at all impurity levels. [ABSTRACT FROM AUTHOR]

Accurate position-force control is a core and challenging problem in robotics, especially for manipulators with redundant degrees of freedom (DOFs). For example, trajectory tracking-based control usually fails for grinding robots due to intolerable impact forces imposed onto the end effectors. The main difficulties lie in the coupling of motion and contact force, redundancy resolution, physical constraints, etc. In this article, we propose a novel motion-force control strategy in the framework of projection recurrent neural networks (RNN). Tracking error and contact force are described in orthogonal spaces, respectively, and by selecting minimizing joint torque as secondary task, the control problem is formulated as a quadratic-programming (QP) problem under multiple constraints. In order to obtain real-time optimization of joint toque, which is nonconvex relative to joint angles, the original QP is reconstructed in the velocity level, where the original objective function is replaced by its time derivative. Then, a dynamic neural network, which is convergence provable is established to solve the modified QP problem online. This work generalizes projection RNN-based position control of manipulators to that of position-force control, which opens a new avenue to shift position-force control of manipulators from pure control perspective to cross design with both convergence and optimality consideration. Numerical and experimental results show that the proposed scheme achieves accurate position-force control, and is capable of handling inequality constraints such as joint angular, velocity, and torque limitations, simultaneously, consumption of joint torque can be decreased effectively. [ABSTRACT FROM AUTHOR]