Performance of Computer Vision Algorithms for Fine‐Grained Classification Using Crowdsourced Insect Images

Deep (machine) learning in recent years has significantly increased the predictive modeling strengths for applications in the areas of computer vision, speech recognition, and natural language processing. Being heavily data-driven, conventional general-purpose deep learning methods mainly rely on training data to learn task-specific data representations. In this thesis, we focus on leveraging novel learning algorithms in deep learning to tackle visual recognition challenges at the three key levels -- namely pixel, object, and image levels. Throughout the thesis, we demonstrate how each and every of our proposed methods can address a specific visual recognition problem. Firstly, we apply deep learning to improve the saliency detection performance of deep networks. A coarse saliency map of the input image's salient object is initially predicted by a convolutional-deconvolutional network (or equivalently a fully convolutional network). Following that, the coarse saliency map is sequentially refined by a separate network called Recurrent Attentional Convolutional-Deconvolutional Network (RACDNN) that shares its network weights across its sequential steps. At every sequential (temporal) step, the same RACDNN is applied to a newly selected/attended image sub-region and it performs saliency map refinement for that particular sub-region. Despite having many learnable network parameters and no access to additional training data, RADCNN can be trained well and it outperforms the "single-step" (coarse saliency detection) baseline and state-of-the-art saliency detection methods by large margins, on several saliency detection benchmark datasets. Secondly, we introduce auto-encoder weight transfer networks (AE-WTN+), a deep model for scaling up object detection. Conventional WTN transfers object class knowledge from the classification weights of a pretrained large-scale image classification network, to the classification weights of an object detection network. However, they suffer from under-fitting and are not able to preserve the rich object class information of novel classes. In this work, we propose a new WTN method -- AE-WTN+ which makes use of normalization techniques to ease network optimization. Inputs and intermediate features are normalized to improve trainability for reduced under-fitting. More importantly, a novel auto-encoding (reconstruction) loss is proposed to encourage information preservation of all classes in the weight predictions of AE-WTN+. This enables the predicted weights of AE-WTN+ to maintain the class discriminability and neighbourhood relationships contained in the pre-trained large-scale classification weights. Experiments on large-scale detection datasets validate the effectiveness of AE-WTN+ on both seen and novel classes. Thirdly, we improve the efficiency of cross-layer network connections for very deep convolutional network in the context of image classification. This gives rise to a new convolutional network architecture called Deluge Networks (DelugeNets) which pass the output (feature activations) of any preceding composite network layer as direct inputs to all succeeding composite layers. For the incoming feature activations of preceding composite layers, the succeeding composite layers first apply rather lightweight cross-layer depthwise convolutional weights on them for feature aggregation. The heavy lifting is carried out by the last convolutional weight layer in each preceding composite layer that performs cross-channel convolutions. The proposed technique enables unimpeded information flow among many layers in very deep DelugeNets with great efficiency. Experiments on small-scale and large-scale datasets show that DelugeNets perform competitively to state-of-the-art deep convolutional networks at lower computational and parameter costs. Fourthly, we propose Stochastic Downsampling (SDPoint), a method to overcome a major limitation suffered by conventional deep convolutional networks -- their computational costs during inference are fixed and cannot be changed. SDPoint downsamples the feature maps at a randomly selected point (layer index) in the network hierarchy during training. The many downsampling configurations known as SDPoint instances each entail a unique downsampling point and downsampling ratio. The same convolutional network weights are shared across the different SDPoint instances that downsample feature maps differently. During inference, the computational cost of a network trained with SDPoint can be instantaneously adjusted to fit a given computational budget, by manually selecting an appropriate SDPoint instance. Additionally, sharing weights between different feature map scales provides significant regularization benefits, making the networks trained with SDPoint less sensitive to scales and more competent. In image classification experiments, SDPoint demonstrates significant cost-accuracy and computational advantages, over independently-trained (non-weight-shared) baselines with various inference costs.

Robust image retrieval by cascading a deep quality assessment network

Predictive analysis on cellular traffic is important for the control and monitoring of wireless networks. Cellular traffic prediction is a challenging problem due to the non- stationarity and dynamic spatial-temporal correlation of the traffic. In this paper, we address the problem of accurate traffic prediction in a base station by proposing a deep neural network called RAConv. Its structure includes residual network, attention mechanism, and deep convolutional network. In the proposed architecture, a deep 3D residual convolutional network (ResConv3D) with three residual blocks are employed to learn the local spatial-temporal features. An attention-aided convolutional long short-term memory network (AConvLSTM) is then used to capture the long-term spatial-temporal dependencies. The use of the attention modules enable the network to focus on the most important spatial-temporal information. We evaluate the performance of the proposed RAConv network using a dataset provided by a Canadian wireless service provider. We consider the traffic prediction on two time scales (i.e., hourly and daily), which exhibit different spatial-temporal dependency patterns. Experimental results show that the proposed RAConv network can achieve accurate prediction under both time scales. Results also show that our proposed network provides a lower root-mean-square error (RMSE) than the conventional ConvLSTM baseline scheme.

Large convolutional networks have achieved impressive classification performances recently. To achieve better performance, convolutional network tends to develop into deeper. However, the increase of network depth causes the linear growth of computational complexity, but cannot bring equivalent increase to the classification accuracy. To alleviate this inconsistence, we propose a cascading approach to accelerate the classification of very deep convolutional neural network. By exploiting the entropy metric to analyze the statistic differences of basic networks between the correctly and mistakenly classified images, we can assign the easily distinguished images to the shallow networks for reducing the computational complexity, and leave the difficultly classified images to the deep networks for maintaining the overall performance. Besides, the proposed cascaded networks can take advantage of the complementarity between different networks, which may boost the classification accuracy compared to the deepest network. We perform the experiments using residual networks of different depths on cifar100 dataset, on the condition of obtaining the similar accuracy to the deepest network, the results show that our cascaded ResNet32-ResNet110 and cascaded ResNet32-ResNet164 can reduce the computation time by 48.6% and 44.3% compared to ResNet110 and ResNet164, respectively. And the cascaded ResNet32-ResNet110-ResNet164 can reduce the computation time by 85.4% compared to the very deep Resnet1001.

This thesis addresses the problem of provident vehicle detection at night: The detection of oncoming vehicles by detecting its light artifacts. In particular, this means detecting a vehicle is oncoming even if it is not directly visible. For this purpose, a field study is conducted to understand human provident detection abilities and to record data. With the recorded data, a dataset is derived that can be used to analyze the performance of computer vision algorithms and to train machine learning approaches to detect vehicles providently. Furthermore, a complete computer vision algorithm is designed and trained on the derived dataset to detect light artifacts caused by the headlights of oncoming vehicles. The performance of this algorithm is compared with the measured performance of humans and an in-production computer vision system to demonstrate that the proposed system detects oncoming vehicles faster than humans and even faster than an in-production computer vision system. In addition, to show the usefulness of the developed algorithm, it is deployed in a test vehicle to use the detected light artifacts to control the glare-free high beam system proactively.

Inferior temporal (IT) cortex in human and nonhuman primates serves visual object recognition. Computational object-vision models, although continually improving, do not yet reach human performance. It is unclear to what extent the internal representations of computational models can explain the IT representation. Here we investigate a wide range of computational model representations (37 in total), testing their categorization performance and their ability to account for the IT representational geometry. The models include well-known neuroscientific object-recognition models (e.g. HMAX, VisNet) along with several models from computer vision (e.g. SIFT, GIST, self-similarity features, and a deep convolutional neural network). We compared the representational dissimilarity matrices (RDMs) of the model representations with the RDMs obtained from human IT (measured with fMRI) and monkey IT (measured with cell recording) for the same set of stimuli (not used in training the models). Better performing models were more similar to IT in that they showed greater clustering of representational patterns by category. In addition, better performing models also more strongly resembled IT in terms of their within-category representational dissimilarities. Representational geometries were significantly correlated between IT and many of the models. However, the categorical clustering observed in IT was largely unexplained by the unsupervised models. The deep convolutional network, which was trained by supervision with over a million category-labeled images, reached the highest categorization performance and also best explained IT, although it did not fully explain the IT data. Combining the features of this model with appropriate weights and adding linear combinations that maximize the margin between animate and inanimate objects and between faces and other objects yielded a representation that fully explained our IT data. Overall, our results suggest that explaining IT requires computational features trained through supervised learning to emphasize the behaviorally important categorical divisions prominently reflected in IT.

According to the National Highway Traffic Safety Administration, one in ten fatal crashes and two in ten injury crashes were reported as distracted driver accidents in the United State during 2014. In an attempt to mitigate these alarming statistics, this paper explores using a dashboard camera along with computer vision and machine learning to automatically detect distracted drivers. We consider a dataset that incorporates drivers engaging in seven different distracting behaviors using left and/or right hands. Traditional handcrafted features paired with a Support Vector Machine classifier are contrasted with deep Convolutional Neural Networks. The traditional features include a blend of Histogram of Oriented Gradients and Scale-Invariant Feature Transform descriptors used to create Bags of Words. The deep convolutional methods use transfer learning on AlexNet, VGG-16, and ResNet-152. The results yield 85% accuracy with ResNet and 82.5% accuracy with VGG-16, which outperformed AlexNet by almost 10%. Replacing the fully connected layers by a Support Vector Machine classifier did not improve the classification accuracy. The traditional features yielded much lower accuracy than the deep convolutional networks.

Object classification has been one of the main tasks in computer vision. With the fast development of deep learning, its performance in image classification and object recognition has presented dramatic improvements. In this paper, we propose a new deep convolutional neural network (CNN) architecture for robust object classification. The proposed model is fused with three traditional CNN approaches, Densenet201, Resnet50, and our proposed residual CNN. The fused network architecture allows parallel processing of the multiple networks for keeping the system sped up. A single shot deep convolution network is trained as an object detector to generate all possible candidates of different object classes. The output of each neural network is representing a single vote that is used in the classification process. 3-to-1 voting criteria are applied in the final classification decision between the candidate object classes. Several experiments were conducted to evaluate the performance of the proposed network. The experimental results show that the proposed approach has better performance than the networks used in the fusion process when they act individually. It also has lower miss rates when compared to several states of art methodologies.

Increasing the level of spacecraft autonomy is essential for broadening the reach of solar system exploration. Computer vision has and will continue to play an important role in increasing autonomy of both spacecraft and Earth-based robotic vehicles. This article addresses progress on computer vision for planetary rovers and landers and has four main parts. First, we review major milestones in the development of computer vision for robotic vehicles over the last four decades. Since research on applications for Earth and space has often been closely intertwined, the review includes elements of both. Second, we summarize the design and performance of computer vision algorithms used on Mars in the NASA/JPL Mars Exploration Rover (MER) mission, which was a major step forward in the use of computer vision in space. These algorithms did stereo vision and visual odometry for rover navigation and feature tracking for horizontal velocity estimation for the landers. Third, we summarize ongoing research to improve vision systems for planetary rovers, which includes various aspects of noise reduction, FPGA implementation, and vision-based slip perception. Finally, we briefly survey other opportunities for computer vision to impact rovers, landers, and orbiters in future solar system exploration missions.

Local binary features for texture classification: Taxonomy and experimental study

This paper proposes a method of accurately assessing image quality without a reference image by using a deep convolutional neural network. Existing training based methods usually utilize a compact set of linear filters for learning features of images captured by different sensors to assess their quality. These methods may not be able to learn the semantic features that are intimately related with the features used in human subject assessment. Observing this drawback, this work proposes training a deep convolutional neural network (CNN) with labelled images for image quality assessment. The ReLU in the CNN allows non-linear transformations for extracting high-level image features, providing a more reliable assessment of image quality than linear filters. To enable the neural network to take images of any arbitrary size as input, the spatial pyramid pooling (SPP) is introduced connecting the top convolutional layer and the fully-connected layer. In addition, the SPP makes the CNN robust to object deformations to a certain extent. The proposed method taking an image as input carries out an end-to-end learning process, and outputs the quality of the image. It is tested on public datasets. Experimental results show that it outperforms existing methods by a large margin and can accurately assess the image quality on images taken by different sensors of varying sizes.

In almost all computer vision applications moving objects detection is the crucial step for information extraction. Shadows and ghosts will often introduce errors that will certainly effect the performance of computer vision algorithms, such as object detection, tracking and scene understanding. This paper studies various methods for shadows and ghost detection and proposes a novel user-aided approach for texture preserving shadows and ghost removal from surveillance video. The proposed algorithm addresses limitations in uneven shadow and ghost boundary processing and umbra recovery. This approach first identifies an initial shadow/ghost boundary by growing a user specified shadow outline on an illumination-sensitive image. Interval-variable pixel intensity sampling is introduced to eliminate anomalies, raised from unequal boundaries. This approach extracts the initial scale field by applying local group intensity spline fittings around the shadow boundary area. Bad intensity samples are substituted by their nearest intensities based on a log-normal probability distribution of fitting errors. Finally, it uses a gradual colour transfer to correct post-processing anomalies such as gamma correction and lossy compression.KeywordsShadow and ghost removaluser-assistedspline fittingsinterval variable

Text recognition in natural scene images is a challenging problem in computer vision. Different than the optical character recognition (OCR), text recognition in natural scene images is more complex due to variations in text size, colors, fonts, orientations, complex backgrounds, occlusion, illuminations and uneven lighting conditions. In this paper, we propose a segmentation-free method based on a deep convolutional recurrent neural network to solve the problem of cursive text recognition, particularly focusing on Urdu text in natural scenes. Compared to the non-cursive scripts, Urdu text recognition is more complex due to variations in the writing styles, several shapes of the same character, connected text, ligature overlapping, stretched, diagonal and condensed text. The proposed model gets a whole word image as an input without pre-segmenting into individual characters, and then transforms into the sequence of the relevant features. Our model is based on three components: a deep convolutional neural network (CNN) with shortcut connections to extract and encode the features, a recurrent neural network (RNN) to decode the convolutional features, and a connectionist temporal classification (CTC) to map the predicted sequences into the target labels. To increase the text recognition accuracy further, we explore deeper CNN architectures like VGG-16, VGG-19, ResNet-18 and ResNet-34 to extract more appropriate Urdu text features, and compare the recognition results. To conduct the experiments, a new large-scale benchmark dataset of cropped Urdu word images in natural scenes is developed. The experimental results show that the proposed deep CRNN network with shortcut connections outperform than other network architectures. The dataset is publicly available and can be downloaded from <uri xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">https://data.mendeley.com/datasets/k5fz57zd9z/1</uri> .

Aiming at the existing malware variants detection method based on deep convolutional neural networks (DCNN) has the problem of large computational resource consumption, a decomposing deep neural network (Dec-DCNN) is proposed for optimization. The computational cost was reduced by using single value decomposition (SVD) to split the pre-trained standard convolution operation into two simpler convolution operations. And the optimized network after decomposition does no need to be retrained, which can reduce the number of parameters and calculations while maintaining the detection accuracy of the pre-trained model. Experimental results show that the Dec-DCNN had a detection time of only 46% of DCNN with an overall detection accuracy of 98.5%. At the same time, compared with the current mainstream malicious variant detection model, Dec-DCNN has a stronger ability to express texture features of malware images.

Object detection systems based on the deep convolutional neural network (CNN) have recently made ground- breaking advances on several object detection benchmarks. While the features learned by these high-capacity neural networks are discriminative for categorization, inaccurate localization is still a major source of error for detection. Building upon high-capacity CNN architectures, we address the localization problem by 1) using a search algorithm based on Bayesian optimization that sequentially proposes candidate regions for an object bounding box, and 2) training the CNN with a structured loss that explicitly penalizes the localization inaccuracy. In experiments, we demonstrated that each of the proposed methods improves the detection performance over the baseline method on PASCAL VOC 2007 and 2012 datasets. Furthermore, two methods are complementary and significantly outperform the previous state-of-the-art when combined.