Bernardo B. Gatto

A deep network model based on subspaces: A novel approach for image classification

In this paper, we propose a novel deep neural network based on learning subspaces and convolutional neural network with applications in image classification. Recently, multistage PCA based filter banks have been successfully adopted in convolutional neural networks architectures in many applications including texture classification, face recognition and scene understanding. These approaches have shown to be powerful, with a straightforward implementation that enables a fast prototyping of efficient image classification systems. However, these architectures employ filters based on PCA, which may not achieve high discriminative features in more complicated computer vision datasets. In order to cope with the aforementioned drawback, we propose a Hybrid Subspace Neural Network (HS-Net). The proposed architecture employs filters from both PCA and discriminative filters banks from more sophisticated subspace methods, therefore achieving more representative and discriminative information. In addition, the use of hybrid architecture enables the use of supervised and unsupervised samples, depending on the application, making the introduced architecture quite attractive in practical terms. Exsperimental results on three publicly available datasets demonstrate the effectiveness and the practicability of the proposed architecture.

Image-Set Matching by Two Dimensional Generalized Mutual Subspace Method

In this paper, we present a novel supervised learning algorithm for object recognition from sets of images, where the sets describe most of the variation in an objects appearance caused by lighting, pose and view angle. In this scenario, generalized mutual subspace method (gMSM) has attracted attention for image-set matching due to its advantages in accuracy and robustness. However, gMSM employs PCA, which has high computational cost contrasting to state-of-art appearance-based methods. To create a faster method, we replace the traditional PCA by 2D-PCA and variants on gMSM framework. In general, 2D-PCA and variants require less memory resource than conventional PCA since its covariance matrix is calculated directly from two-dimensional matrices. The introduced method has the advantage of representing the subspaces in a more compact manner, providing reasonably competitive recognition rate comparing to the traditional MSM, confirming the suitability of employing 2D-PCA and variants on gMSM framework. These results have been revealed through experimentation conducted on five widely used datasets.

Enhancing discriminability of randomized time warping for motion recognition

In this paper, we propose a framework of action sequence recognition by combining the representation of randomized time warping (RTW) with the enhanced Grassmann discriminant Analysis (eGDA). RTW is an extension of Dynamic time warping (DTW), and it has been shown to be effective for motion recognition, as it can effectively retain an actions temporal information by generating a low-dimensional subspace from a set of time elastic (TE) features of a video. On the other hand, the eGDA can use the concepts of generalized difference subspace and Grassmann manifold symbiotically to learn a discriminative manifold where video subspaces can be regarded as points. The main advantages of the proposed method are: removing common features between the actions which are not useful for discrimination, thus increasing the distance between subspaces of different classes, and reducing the distance between subspaces of the same class; and estimating a discriminative manifold even if there are few training data. We demonstrate the validity of the proposed method through experiments on motion recognition using two public datasets, namely, the Cambridge gesture database and the KTH action dataset.

Orthogonal Hankel Subspaces for Applications in Gesture Recognition

Gesture recognition is an important research area in video analysis and computer vision. Gesture recognition systems include several advantages, such as the interaction with machines without needing additional external devices. Moreover, gesture recognition involves many challenges, as the distribution of a specific gesture largely varies depending on viewpoints due to its multiple joint structures. In this paper, We present a novel framework for gesture recognition. The novelty of the proposed framework lies in three aspects: first, we propose a new gesture representation based on a compact trajectory matrix, which preserves spatial and temporal information. We understand that not all images of a gesture video are useful for the recognition task, therefore it is necessary to create a method where it is possible to detect the images that do not contribute to the recognition task, decreasing the computational cost of the overall framework. Second, we represent this compact trajectory matrix as a subspace, achieving discriminative information, as the trajectory matrices obtained from different gestures generate dissimilar clusters in a low dimension space. Finally, we introduce an automatic procedure to infer the optimal dimension of each gesture subspace. We show that our compact representation presents practical and theoretical advantages, such as compact representation and low computational requirements. We demonstrate the advantages of the proposed method by experimentation employing Cambridge gesture and Human-Computer Interaction datasets.

A Framework for Chainsaw Detection Using One-Class Kernel and Wireless Acoustic Sensor Networks into the Amazon Rainforest

In this work we present a framework for automatic acoustic detection of chainsaw sounds to detect illegal wood extraction in the Amazon Rainforest. Our approach was developed to be embedded into the sensor nodes of a Wireless Acoustic Sensor Network (WASN) to monitor the environment. First, we represent each sound by a set of Mel-Frequency Cepstral Coefficients (MFCCs) and, after that, we fit a Probability Density Function (PDF) with a kernel based on a multivariate Gaussian density estimation using only the target class. This One-Class classification method allows us to recognize only chainsaw sounds rejecting all the other possible environmental sounds, such as: animals calls, weather noises or boat engines. In the experiments, we varied the number MFCCs coefficients and the Kernel bandwidth performing a leave-one-out cross validation to find the best combination. Finally, we found that the best parameter combination achieve 98% of accuracy showing a low FNR and a high TPR, fact that enhances the credibility of the system avoiding false alarms and making it an optimal choice for an WASN application.