Search

In the last two decades, hyperspectral images have been studied for use in food quality analysis using neural network-based chemometric techniques such as radial basic neural network (RBNN) in classification tasks or multilayer perceptron (MLP) for prediction ones. However, recently, deep learning techniques are receiving special interest as those based on recurrent neural networks (RNN) or gated recurrent units for their feature extraction capacity. A kind of RNN are networks based on long short-term memory (LSTM), which are promising for improving the capabilities of hyperspectral images in food analysis. In this sense, a comparative study on using LSTM-based networks is done, comparing it to RBNN in the discrimination of fish fillets subjected to freeze-thaw cycles. The samples for the study were fresh mackerel fillets subjected to two successive cycles of freezing and thawing. Hyperspectral images of fillets were acquired, both fresh and in each freeze-thaw cycle, and spectral profiles were acquired in the range of 400–1000 nm. Complete spectral profiles were used to generate classification models; these were subsequently optimized using the relevant wavelengths determined by feature selection neighborhood component analysis. The models were trained using a k-cross-fold validation strategy (k = 5), repeating the process thirty times. The models were evaluated, confusion matrices were determined, and their classification metrics were compared. The training process results showed that the LSTM networks had superior potential to the RBNN networks, achieving an average F-measure greater than 0.89 and an accuracy greater than 0.93. This study shows that deep learning techniques can be used to tell the difference between fish fillets that have been frozen and thawed several times. These techniques could also be used to establish systems to control the transport and conservation processes of this type of food. [ABSTRACT FROM AUTHOR]

A correction to the article "Supercomputing and grid computing on the verification of covering arrays," published in a previous issue is presented.

The Covering Arrays (CAs) are mathematical objects with minimal coverage and maximum cardinality that are a good tool for the design of experiments. A covering array is an N× k matrix over an alphabet v s.t. each N× k subset contains at least one time each combination from {0,1,..., v−1}, given a positive integer value t. The process of ensuring that a CA contains each of the v combinations is called verification of CA. In this paper, we present an algorithm for CA verification and its implementation details in three different computation paradigms: (a) sequential approach (SA); (b) parallel approach (PA); and (c) Grid approach (GA). Four different PAs were compared in their performance of verifying a matrix as a CA; the PA with the best performance was included in a different experimentation where the three paradigms, SA, PA, and GA were compared in a benchmark composed by 45 possible CA instances. The results showed the limitations of the different paradigms when solving the verification of CA problem, and points out the necessity of a Grid approach to solve the problem when the size of a CA grows. [ABSTRACT FROM AUTHOR]