Article Preview
TopIntroduction
In recent years, there has been significant interest in the area of video analysis. The practical applications of such systems could include video surveillance for military use, or monitoring of activities in secure installations, action recognition, sports and pattern recognition, amongst others. Dictionary learning (DL) is an effective approach for video semantic content recognition, originally derived from compressed sensing theory as discussed in (Liu et al., 2020) and has also made some remarkable achievements for video semantic analysis (Benuwa et al., 2019; Boccignone et al., 2017; Liu et al., 2020). Dictionaries sometimes are constructed using atoms from a predefined set, which may be simple but not an appropriate approach, particularly in cases where the task is more complicated (Abdi et al., 2019; Rubinstein et al., 2010). The dictionary could also be learned in an optimization process on the other hand, in a way that the dictionary atoms will fit the training samples with the appropriate properties (such as discrimination, reconstruction, de-noising etc.) indicated in the dictionary. These various DL tasks have been handled by a lot of researchers and have shown impressive outcomes in many complex situations as discussed in (Abdi et al., 2019; Benuwa et al., 2018; Gao et al., 2013). DL from an angle, could be categorized into discriminative and reconstructive depending on the properties assigned the learned dictionary. A lot of Discriminative based DL (DDL) approaches proposed for classification tasks use sparse representation and incorporate discriminability into the structure of the DL process (Abdi et al., 2019; Akhtar et al., 2016; Gao et al., 2013; Liu et al., 2015). In some other studies, the dictionary is segregated into class specific sub-dictionaries and the samples belonging to each class enforced to use their corresponding sub-dictionary entities in the decomposition. In making the dictionary discriminative, it is ideal to introduce a discriminative parameter as a term in the optimization process as have been reported in (Abdi et al., 2019) with the addition of cost function (Golmohammady et al., 2014; Yang et al., 2011), classification error (Jiang et al., 2013; Qiu et al., 2014; Zhang & Li, 2010) and information theoretic (Lazebnik & Raginsky, 2008; Qiu et al., 2014).
Conversely, majority of reconstructive DL techniques are aimed at realizing minimized reconstruction errors with a maximized sparsity of the representation coefficient during the optimization process. Some of the existing approaches that employed this technique can be found in (Aharon et al., 2006; Engan et al., 1999; Liu et al., 2020). In cases where the samples are from different categories, these algorithms could be used for classification despite having being design for reconstruction tasks. Though incorporating discriminative property into the structure or the objective functions of these methods (to enhance the discriminability of the learned dictionaries), it is worth noting that, this may decrease the reconstruction capabilities of the dictionary simultaneously (Abdi et al., 2019; Liu et al., 2020). Therefore, there should be a trade-off criterion among discrimination and reconstruction criterions of these DL approaches with respect to their objective functions. This may respectively lead to generalization ability of the learned dictionary leading to appropriate classification accuracy occurring in the test sample as a result of the reconstruction term. On the other hand, a better classification result is achieved in the training samples on the other hand due to the discriminative ability of the dictionary as a result of the discriminative term.