Visual object recognition and detection using deep learning

Visual object recognition problem is defined as the determination of the object inside an image file. With the advent of Deep Learning, learning a good, lower dimensional representation for the raw data is made possible with the use of deep architectures. Deep Belief Networks are proposed as a multi-layered probabilistic model where each layer is trained by maximizing the probability of the previous layer's output over the model distribution. Said approach corresponds to a greedy layer-wise unsupervised training procedure where the object representations are learned so that the data probability over the model is maximized. In the first part of the thesis, we worked on a sample object recognition application, specifically handwritten digit recognition. Handwritten digit recognition is a problem first emerged from the need to correctly recognize ZIP codes on the letters posted through U.S. Postal Service. While the recognition of single characters is a fairly well studied research field, it is an important part of any system used for generating text from an image containing handwriting. MNIST dataset is a specialized dataset collected for the task of handwritten digit recognition. Because MNIST provides researchers with a simple and well defined problem, it is widely used as a benchmark for Machine Learning algorithms. Overfitting of the model to the training data is also a problem of deep architectures as well as classical Machine Learning approaches. To prevent overfitting and generate generalizable models, regularization is introduced to the optimization procedure in the training. Regularization is used as a powerfull tool to avoid overfitting the complex structures in the training data which might not be present in unobserved data samples. Regularization can also be seen as a tool to impose a prior on the representation model in a Bayesian sense. We propose a Clique Potential regularization which is form of morphological constraints on the connections between a Restricted Boltzmann Machine. If the input data is an image, the connections between layers can also be viewed as images. Clique Potential regularization puts costs to not connected neighbours so that the features learned will be sparse and noise free. Introducing a regularization to the model also presents us with the challenge of determining a hyperparameter for said regularization. Selecting a hyperparameter is also defined as an optimization problem where evaluating the objective function is computationally costly, because it requires a full training cycle to observe the effects of the hyperparameter selection. We employ Bayesian optimization to select the hyperparameter of the regularization. Bayesian optimization methods are used in literature to optimize costly black-box objective functions with minimum number of evaluations, as they model the objective function and decide the next point to evaluate based on the expected gain of the input space. The hyperparameters of Bayesian optimization is marginalized to procure a parameter-free method. In the first part of the thesis, performance of the system is evaluated based on the classification error over the test data. Deep Belief Network (DBN) architecture with no regularizations achieves 1.57\% test error, while DBN with a 4 neighbourhood Clique Potential regularization gives the result of 1.49\% test error. 4 neighbourhood Clique Potential regularization results with features that is noise free and more sparse. $l_1$ and $l_2$ regularizations both generate 1.46\% classification error on the test set. $l_1$ regularization pushes most of the weights towards zero, therefore generate a more sparse representation. $l_2$ regularization punishes the weights with higher absolute values more aggressively, so the noise in the uniform gray level regions is not effected by $l_2$. Visual object detection is the problem of determining the location and the type of, possibly multiple, objects inside an image. For the second part of this thesis, we worked on Faster R-CNN method, which is the current state-of-the-art method for visual object detection, to generate a detection model for VOT2016 object tracking challenge where different objects required to be tracked in different videos. A custom dataset is created by using publicly available datasets to cover classes required for VOT2016 challenge. Overall system performance is measured by using mean average precision metric employed by COCO challenge. %Overall system performance was observed as 27\% according to mean average precision metric defined by COCO dataset, and 48.6\% according to average precision metric used by Pascal VOC dataset.