Search

Biotic and abiotic plant stress (e.g., frost, fungi, diseases) can significantly impact crop production. It is thus essential to detect such stress at an early stage before visual symptoms and damage become apparent. To this end, this paper proposes a novel deep learning method, called Spectral Convolution and Channel Attention Network (SC-CAN), which exploits the difference in spectral responses of healthy and stressed crops. The proposed SC-CAN method comprises two main modules: (i) a spectral convolution module, which consists of dilated causal convolutional layers stacked in a residual manner to capture the spectral features; (ii) a channel attention module, which consists of a global pooling layer and fully connected layers that compute inter-relationship between feature map channels before scaling them based on their importance level (attention score). Unlike standard convolution, which focuses on learning local features, the dilated convolution layers can learn both local and global features. These layers also have long receptive fields, making them suitable for capturing long dependency patterns in hyperspectral data. However, because not all feature maps produced by the dilated convolutional layers are important, we propose a channel attention module that weights the feature maps according to their importance level. We used SC-CAN to classify salt stress (i.e., abiotic stress) on four datasets (Chinese Spring (CS), Aegilops columnaris (co(CS)), Ae. speltoides auchery (sp(CS)), and Kharchia datasets) and Fusarium head blight disease (i.e., biotic stress) on Fusarium dataset. Reported experimental results show that the proposed method outperforms existing state-of-the-art techniques with an overall accuracy of 83.08%, 88.90%, 82.44%, 82.10%, and 82.78% on CS, co(CS), sp(CS), Kharchia, and Fusarium datasets, respectively.

Modern deep learning methods constitute incredibly powerful tools to tackle a myriad of challenging problems. However, since deep learning methods operate as black boxes, the uncertainty associated with their predictions is often challenging to quantify. Bayesian statistics offer a formalism to understand and quantify the uncertainty associated with deep neural network predictions. This tutorial provides deep learning practitioners with an overview of the relevant literature and a complete toolset to design, implement, train, use and evaluate Bayesian neural networks, i . e ., stochastic artificial neural networks trained using Bayesian methods.

Weakly supervised semantic segmentation has been attracting increasing attention as it can alleviate the need for expensive pixel-level annotations through the use of image-level labels. Relevant methods mainly rely on the implicit object localization ability of convolutional neural networks (CNNs). However, generated object attention maps remain mostly small and incomplete. In this paper, we propose an Atrous Convolutional Feature Network (ACFN) to generate dense object attention maps. This is achieved by enhancing the context representation of image classification CNNs. More specifically, cascaded atrous convolutions are used in the middle layers to retain sufficient spatial details, and pyramidal atrous convolutions are used in the last convolutional layers to provide multi-scale context information for the extraction of object attention maps. Moreover, we propose an attentive fusion strategy to adaptively fuse the multi-scale features. Our method shows improvements over existing methods on both the PASCAL VOC 2012 and MS COCO datasets, achieving state-of-the-art performance.

Semantic segmentation is a challenging task in the absence of densely labelled data. Only relying on class activation maps (CAM) with image-level labels provides deficient segmentation supervision. Prior works thus consider pre-trained models to produce coarse saliency maps to guide the generation of pseudo segmentation labels. However, the commonly used off-line heuristic generation process cannot fully exploit the benefits of these coarse saliency maps. Motivated by the significant inter-task correlation, we propose a novel weakly supervised multi-task framework termed as AuxSegNet, to leverage saliency detection and multi-label image classification as auxiliary tasks to improve the primary task of semantic segmentation using only image-level ground-truth labels. Inspired by their similar structured semantics, we also propose to learn a cross-task global pixellevel affinity map from the saliency and segmentation representations. The learned cross-task affinity can be used to refine saliency predictions and propagate CAM maps to provide improved pseudo labels for both tasks. The mutual boost between pseudo label updating and cross-task affinity learning enables iterative improvements on segmentation performance. Extensive experiments demonstrate the effectiveness of the proposed auxiliary learning network structure and the cross-task affinity learning method. The proposed approach achieves state-of-the-art weakly supervised segmentation performance on the challenging PASCAL VOC 2012 and MS COCO benchmarks.

In this paper, we propose a high performance Two-Stream spectral-spatial Residual Network (TSRN) for hyperspectral image classification. The first spectral residual network (sRN) stream is used to extract spectral characteristics, and the second spatial residual network (saRN) stream is concurrently used to extract spatial features. The sRN uses 1D convolutional layers to fit the spectral data structure, while the saRN uses 2D convolutional layers to match the hyperspectral spatial data structure. Furthermore, each convolutional layer is preceded by a Batch Normalization (BN) layer that works as a regularizer to speed up the training process and to improve the accuracy. We conducted experiments on three well-known hyperspectral datasets, and we compare our results with five contemporary methods across various sizes of training samples. The experimental results show that the proposed architecture can be trained with small size datasets and outperforms the state-of-the-art methods in terms of the Overall Accuracy, Average Accuracy, Kappa Value, and training time.

Weakly supervised semantic segmentation with image-level labels is of great significance since it alleviates the dependency on dense annotations. However, as it relies on image classification networks that are only capable of producing sparse object localization maps, its performance is far behind that of fully supervised semantic segmentation models. Inspired by the successful use of multi-scale features for an improved performance in a wide range of visual tasks, we propose a Scale-Aware Feature Network (SAFN) for generating object localization maps. The proposed SAFN uses an attention module to learn the relative weights of multi-scale features in a modified fully convolutional network with dilated convolutions. This approach leads to efficient enlargements of the receptive fields of view and produces dense object localization maps. Our approach achieves mIoUs of 62.3% and 66.5% on the PASCAL VOC 2012 test set using VGG16 based and ResNet based segmentation models, respectively, outperforming other state-of-the-art methods for the weakly supervised semantic segmentation task.

Oceanographers rely on advanced digital imaging systems to assess the health of marine ecosystems. The majority of the imagery collected by these systems do not get annotated due to lack of resources. Consequently, the expert labeled data is not enough to train dedicated deep networks. Meanwhile, in the deep learning community, much focus is on how to use pre-trained deep networks to classify out-of-domain images and transfer learning. In this paper, we leverage these advances to evaluate how well features extracted from deep neural networks transfer to underwater image classification. We propose new image features (called ResFeats) extracted from the different convolutional layers of a deep residual network pre-trained on ImageNet. We further combine the ResFeats extracted from different layers to obtain compact and powerful deep features. Moreover, we show that ResFeats consistently perform better than their CNN counterparts. Experimental results are provided to show the effectiveness of ResFeats with state-of-the-art classification accuracies on MLC, Benthoz15, EILAT and RSMAS datasets.

Across the globe, remote image data is rapidly being collected for the assessment of benthic communities from shallow to extremely deep waters on continental slopes to the abyssal seas. Exploiting this data is presently limited by the time it takes for experts to identify organisms found in these images. With this limitation in mind, a large effort has been made globally to introduce automation and machine learning algorithms to accelerate both classification and assessment of marine benthic biota. One major issue lies with organisms that move with swell and currents, such as kelps. This paper presents an automatic hierarchical classification method local binary classification as opposed to the conventional flat classification to classify kelps in images collected by autonomous underwater vehicles. The proposed kelp classification approach exploits learned feature representations extracted from deep residual networks. We show that these generic features outperform the traditional off-the-shelf CNN features and the conventional hand-crafted features. Experiments also demonstrate that the hierarchical classification method outperforms the traditional parallel multi-class classifications by a significant margin (90.0% vs. 57.6% and 77.2% vs. 59.0%) on Benthoz15 and Rottnest datasets respectively. Furthermore, we compare different hierarchical classification approaches and experimentally show that the sibling hierarchical training approach outperforms the inclusive hierarchical approach by a significant margin. We also report an application of our proposed method to study the change in kelp cover over time for annually repeated AUV surveys.

Estimating depth from RGB images is a long-standing ill-posed problem, which has been explored for decades by the computer vision, graphics, and machine learning communities. Among the existing techniques, stereo matching remains one of the most widely used in the literature due to its strong connection to the human binocular system. Traditionally, stereo-based depth estimation has been addressed through matching hand-crafted features across multiple images. Despite the extensive amount of research, these traditional techniques still suffer in the presence of highly textured areas, large uniform regions, and occlusions. Motivated by their growing success in solving various 2D and 3D vision problems, deep learning for stereo-based depth estimation has attracted a growing interest from the community, with more than 150 papers published in this area between 2014 and 2019. This new generation of methods has demonstrated a significant leap in performance, enabling applications such as autonomous driving and augmented reality. In this paper, we provide a comprehensive survey of this new and continuously growing field of research, summarize the most commonly used pipelines, and discuss their benefits and limitations. In retrospect of what has been achieved so far, we also conjecture what the future may hold for deep learning-based stereo for depth estimation research.

Human actions represented with 3D skeleton sequences are robust to clustered backgrounds and illumination changes. In this paper, we investigate skeleton-based action prediction, which aims to recognize an action from a partial skeleton sequence that contains incomplete action information. We propose a new Latent Global Network based on adversarial learning for action prediction. We demonstrate that the proposed network provides latent long-term global information that is complementary to the local action information of the partial sequences and helps improve action prediction. We show that action prediction can be improved by combining the latent global information with the local action information. We test the proposed method on three challenging skeleton datasets and report state-of-the-art performance.

Underwater imaging is being extensively used for monitoring the abundance of lobster species and their biodiversity in their local habitats. However, manual assessment of these images requires a huge amount of human effort. In this article, we propose to automate the process of lobster detection using a deep learning technique. A major obstacle in deploying such an automatic framework for the localization of lobsters in diverse environments is the lack of large annotated training datasets. Generating synthetic datasets to train these object detection models has become a popular approach. However, the current synthetic data generation frameworks rely on automatic segmentation of objects of interest, which becomes difficult when the objects have a complex shape, such as lobster. To overcome this limitation, we propose an approach to synthetically generate parts of the lobster. To handle the variability of real-world images, these parts were inserted into a set of diverse background marine images to generate a large synthetic dataset. A state-of-the-art object detector was trained using this synthetic parts dataset and tested on the challenging task of Western rock lobster detection in West Australian seas. To the best of our knowledge, this is the first automatic lobster detection technique for partially visible and occluded lobsters.

Human actions represented with 3D skeleton sequences are robust to clustered backgrounds and illumination changes. In this paper, we investigate skeleton-based action prediction, which aims to recognize an action from a partial skeleton sequence that contains incomplete action information. We propose a new Latent Global Network based on adversarial learning for action prediction. We demonstrate that the proposed network provides latent long-term global information that is complementary to the local action information of the partial sequences and helps improve action prediction. We show that action prediction can be improved by combining the latent global information with the local action information. We test the proposed method on three challenging skeleton datasets and report state-of-the-art performance.

In this paper, we propose a new approach to recognize the class label of an action before this action is fully performed based on skeleton sequences. Compared to action recognition which uses fully observed action sequences, early action recognition with partial sequences is much more challenging mainly due to: (1) the global information of a long-term action is not available in the partial sequence, and (2) the partial sequences at different observation ratios of an action contain a number of sub-actions with diverse motion information. To address the first challenge, we introduce a global regularizer to learn a hidden feature space, where the statistical properties of the partial sequences are similar to those of the full sequences. We introduce a temporal-aware cross-entropy to address the second challenge and achieve better prediction performance. We evaluate the proposed method on three challenging skeleton datasets. Experimental results show the superiority of the proposed method for skeleton-based early action recognition.

Coral classification is a challenging task due to the complex morphology and ambiguous boundaries of corals. This paper investigates the benefits of Densely connected convolutional network (DenseNet) and multi-modal image translation techniques in boosting image classification performance by synthesizing missing fluorescence information. To this end, an imageconditional Generative Adversarial Network (GAN) based image translator is trained to model the relationship between reflectance and fluorescence images. Through this image translator, fluorescence images can be generated from the available reflectance images to provide complementary information. During the classification phase, reflectance and translated fluorescence images are combined to obtain more discriminative representations and produce improved classification performance. We present results on the EFC and MLC datasets and report state-of-the-art coral classification performance.

Weakly supervised semantic segmentation with image-level labels is of great significance since it alleviates the dependency on dense annotations. However, it is a challenging task as it aims to achieve a mapping from high-level semantics to low-level features. In this work, we propose a three-step method to bridge this gap. First, we rely on the interpretable ability of deep neural networks to generate attention maps with class localization information by back-propagating gradients. Secondly, we employ an off-the-shelf object saliency detector with an iterative erasing strategy to obtain saliency maps with spatial extent information of objects. Finally, we combine these two complementary maps to generate pseudo ground-truth images for the training of the segmentation network. With the help of the pre-trained model on the MS-COCO dataset and a multi-scale fusion method, we obtained mIoU of 62.1% and 63.3% on PASCAL VOC 2012 val and test sets, respectively, achieving new state-of-the-art results for the weakly supervised semantic segmentation task.

Research on marine species recognition is an important part of the actions for the protection of the ocean environment. It is also an under-exploited application area in the computer vision community. However, with the developments of deep learning, there has been an increasing interest about this topic. In this chapter, we present a comprehensive review of the computer vision techniques for marine species recognition, mainly from the perspectives of both classification and detection. In particular, we focus on capturing the evolution of various deep learning techniques in this area. We further compare the contemporary deep learning techniques with traditional machine learning techniques, and discuss the complementary issues between these two approaches. This chapter examines the attributes and challenges of a number of popular marine species datasets (which involve coral, kelp, plankton and fish) on recognition tasks. In the end, we highlight a few potential future application areas of deep learning in marine image analysis such as segmentation and enhancement of image quality.

3-D object recognition is a challenging task for many applications including autonomous robot navigation and scene understanding. Accurate recognition relies on the selection/learning of discriminative features that are in turn used to uniquely characterize the objects. This paper proposes a novel evolutionary feature learning (EFL) technique for 3-D object recognition. The proposed novel automatic feature learning approach can operate directly on 3-D raw data, alleviating the need for data pre-processing, human expertise and/or defining a large set of parameters. EFL offers smart search strategy to learn the best features in a huge feature space to achieve superior recognition performance. The proposed technique has been extensively evaluated for the task of 3-D object recognition on four popular data sets including Washington RGB-D (low resolution 3-D Video), CIN 2D3D, Willow 2D3D and ETH-80 object data set. Reported experimental results and evaluation against existing state-of-the-art methods (e.g., unsupervised dictionary learning and deep networks) show that the proposed EFL consistently achieves superior performance on all these data sets.

Person re-identification (re-ID), which aims to identify the same individual from a gallery collected with different cameras, has attracted increasing attention in the multimedia retrieval community. Current deep learning methods for person re-identification (re-ID) focus on learning classification models on training identities to obtain a ID-discriminative Embedding (IDE) extractor, which is used to extract features from testing images for re-ID. The IDE features of the testing identities might not be discriminative due to that the training identities are different from the testing identities. In this paper, we introduce a new ID-Adaptation Network (ID-AdaptNet), which aims to improve the discriminative power of the IDE features of the testing identities for better person re-ID. The main idea of the ID-AdaptNet is to transform the IDE features to a common discriminative latent space, where the representations of the ‘seen’ training identities are enforced to adapt to those of the ‘unseen’ training identities. More specifically, the ID-AdaptNet is trained by simultaneously minimizing the classification cross-entropy and the discrepancy between the ‘seen’ and the ‘unseen’ training identities in the hidden space. To calculate the discrepancy, we represent their probability distributions as moment sequences and calculate their distance using their central moments. We further propose a Stacking ID-AdaptNet that jointly trains multiple ID-AdaptNets with a regularization method for better re-ID. Experiments show that the ID-AdaptNet and Stacking ID-AdaptNet effectively improve the discriminative power of IDE features.

By introducing sign constraints on the weights, this paper proposes sign constrained rectifier networks (SCRNs), whose training can be solved efficiently by the well known majorization–minimization (MM) algorithms. We prove that the proposed two-hidden-layer SCRNs, which exhibit negative weights in the second hidden layer and negative weights in the output layer, are capable of separating any number of disjoint pattern sets. Furthermore, the proposed two-hidden-layer SCRNs can decompose the patterns of each class into several clusters so that each cluster is convexly separable from all the patterns from the other classes. This provides a means to learn the pattern structures and analyse the discriminant factors between different classes of patterns. Experimental results are provided to show the benefits of sign constraints in improving classification performance and the efficiency of the proposed MM algorithm.

This paper presents a new representation of skeleton sequences for 3D action recognition. Existing methods based on hand-crafted features or recurrent neural networks cannot adequately capture the complex spatial structures and the long-term temporal dynamics of the skeleton sequences, which are very important to recognize the actions. In this paper, we propose to transform each channel of the 3D coordinates of a skeleton sequence into a clip. Each frame of the generated clip represents the temporal information of the entire skeleton sequence, and one particular spatial relationship between the skeleton joints. The entire clip incorporates multiple frames with different spatial relationships, which provide useful spatial structural information of the human skeleton. We also propose a Multi-task Convolutional Neural Network (MTCNN) to learn the generated clips for action recognition. The proposed MTCNN processes all the frames of the generated clips in parallel to explore the spatial and temporal information of the skeleton sequences. The proposed method has been extensively tested on six challenging benchmark datasets. Experimental results consistently demonstrate the superiority of the proposed clip representation and feature learning method for 3D action recognition compared to existing techniques.

Healthy coral reefs play a vital role in maintaining biodiversity in tropical marine ecosystems. Remote imaging techniques have facilitated the scientific investigations of these intricate ecosystems, particularly at depths beyond 10 m where SCUBA diving techniques are not time or cost efficient. With millions of digital images of the seafloor collected using remotely operated vehicles and autonomous underwater vehicles (AUVs), manual annotation of these data by marine experts is a tedious, repetitive, and time-consuming task. It takes 10–30 min for a marine expert to meticulously annotate a single image. Automated technology to monitor the health of the oceans would allow for transformational ecological outcomes by standardizing methods to detect and identify species. This paper aims to automate the analysis of large available AUV imagery by developing advanced deep learning tools for rapid and large-scale automatic annotation of marine coral species. Such an automated technology would greatly benefit marine ecological studies in terms of cost, speed, and accuracy. To this end, we propose a deep learning based classification method for coral reefs and report the application of the proposed technique to the automatic annotation of unlabeled mosaics of the coral reef in the Abrolhos Islands, W.A., Australia. Our proposed method automatically quantified the coral coverage in this region and detected a decreasing trend in coral population, which is in line with conclusions drawn by marine ecologists.

© 2018 IEEE. Coral species, with complex morphology and ambiguous boundaries, pose a great challenge for automated classification. CNN activations, which are extracted from fully connected layers of deep networks (FC features), have been successfully used as powerful universal representations in many visual tasks. In this paper, we investigate the transferability and combined performance of FC features and CONY features (extracted from convolutional layers) in the coral classification of two image modalities (reflectance and fluorescence), using a typical deep network (e.g. VGGNet). We exploit vector of locally aggregated descriptors (VLAD) encoding and principal component analysis (PCA) to compress dense CONY features into a compact representation. Experimental results demonstrate that encoded CONV3 features achieve superior performances on reflectance and fluorescence coral images, compared to FC features. The combination of these two features further improves the overall accuracy and achieves state-of-the-art performance on the challenging EFC dataset.

Predicting an interaction before it is fully executed is very important in applications, such as human-robot interaction and video surveillance. In a two-human interaction scenario, there are often contextual dependency structures between the global interaction context of the two humans and the local context of the different body parts of each human. In this paper, we propose to learn the structure of the interaction contexts and combine it with the spatial and temporal information of a video sequence to better predict the interaction class. The structural models, including the spatial and the temporal models, are learned with long short term memory (LSTM) networks to capture the dependency of the global and local contexts of each RGB frame and each optical flow image, respectively. LSTM networks are also capable of detecting the key information from global and local interaction contexts. Moreover, to effectively combine the structural models with the spatial and temporal models for interaction prediction, a ranking score fusion method is introduced to automatically compute the optimal weight of each model for score fusion. Experimental results on the BIT-Interaction Dataset and the UT-Interaction Dataset clearly demonstrate the benefits of the proposed method.

© 2018 Elsevier Ltd By introducing sign constraints on the weights, this paper proposes sign constrained rectifier networks (SCRNs), whose training can be solved efficiently by the well known majorization–minimization (MM) algorithms. We prove that the proposed two-hidden-layer SCRNs, which exhibit negative weights in the second hidden layer and negative weights in the output layer, are capable of separating any number of disjoint pattern sets. Furthermore, the proposed two-hidden-layer SCRNs can decompose the patterns of each class into several clusters so that each cluster is convexly separable from all the patterns from the other classes. This provides a means to learn the pattern structures and analyse the discriminant factors between different classes of patterns. Experimental results are provided to show the benefits of sign constraints in improving classification performance and the efficiency of the proposed MM algorithm.

The three-dimensional (3D) modeling and recognition of 3D objects have been traditionally performed using local features to represent the underlying 3D surface. Extraction of features requires cropping of several local surface patches around detected keypoints. Although an important step, the extraction and representation of such local patches adds to the computational complexity of the algorithms. This paper proposes a novel Keypoints-based Surface Representation (KSR) technique. The proposed technique has the following two characteristics: (1) It does not rely on the computation of features on a small surface patch cropped around a detected keypoint. Rather, it exploits the geometrical relationship between the detected 3D keypoints for local surface representation. (2) KSR is computationally efficient, requiring only seconds to process 3D models with over 50,000 points with a MATLAB implementation. Experimental results on the UWA and Stanford 3D models dataset suggest that it can accurately perform pairwise and multiview range image registration (3D modeling). KSR was also tested for 3D object recognition with occluded scenes. Recognition results on the UWA dataset show that the proposed technique outperforms existing methods including 3D-Tensor, VD-LSD, keypoint-depth based feature, spherical harmonics and spin image with a recognition rate of 95.9%. The proposed approach also achieves a recognition rate of 93.5% on the challenging Ca'Fascori dataset compared to 92.5% achieved by game-theoretic. The proposed method is computationally efficient compared to state-of-the-art local feature methods.

Predicting an interaction before it is fully executed is very important in applications such as human-robot interaction and video surveillance. In a two-human interaction scenario, there often contextual dependency structure between the global interaction context of the two humans and the local context of the different body parts of each human. In this paper, we propose to learn the structure of the interaction contexts, and combine it with the spatial and temporal information of a video sequence for a better prediction of the interaction class. The structural models, including the spatial and the temporal models, are learned with Long Short Term Memory (LSTM) networks to capture the dependency of the global and local contexts of each RGB frame and each optical flow image, respectively. LSTM networks are also capable of detecting the key information from the global and local interaction contexts. Moreover, to effectively combine the structural models with the spatial and temporal models for interaction prediction, a ranking score fusion method is also introduced to automatically compute the optimal weight of each model for score fusion. Experimental results on the BIT-Interaction and the UT-Interaction datasets clearly demonstrate the benefits of the proposed method.

This letter presents SkeletonNet, a deep learning framework for skeleton-based 3-D action recognition. Given a skeleton sequence, the spatial structure of the skeleton joints in each frame and the temporal information between multiple frames are two important factors for action recognition. We first extract body-part-based features from each frame of the skeleton sequence. Compared to the original coordinates of the skeleton joints, the proposed features are translation, rotation, and scale invariant. To learn robust temporal information, instead of treating the features of all frames as a time series, we transform the features into images and feed them to the proposed deep learning network, which contains two parts: one to extract general features from the input images, while the other to generate a discriminative and compact representation for action recognition. The proposed method is tested on the SBU kinect interaction dataset, the CMU dataset, and the large-scale NTU RGB+D dataset and achieves state-of-the-art performance

This paper presents a new method for 3D action recognition with skeleton sequences (i.e., 3D trajectories of human skeleton joints). The proposed method first transforms each skeleton sequence into three clips each consisting of several frames for spatial temporal feature learning using deep neural networks. Each clip is generated from one channel of the cylindrical coordinates of the skeleton sequence. Each frame of the generated clips represents the temporal information of the entire skeleton sequence, and incorporates one particular spatial relationship between the joints. The entire clips include multiple frames with different spatial relationships, which provide useful spatial structural information of the human skeleton. We propose to use deep convolutional neural networks to learn long-term temporal information of the skeleton sequence from the frames of the generated clips, and then use a Multi-Task Learning Network (MTLN) to jointly process all frames of the generated clips in parallel to incorporate spatial structural information for action recognition. Experimental results clearly show the effectiveness of the proposed new representation and feature learning method for 3D action recognition.

This chapter presents a summary of the use of deep learning for underwater image analysis, in particular for coral species classification. Deep learning techniques have achieved the state-of-the-art results in various computer vision tasks such as image classification, object detection, and scene understanding. Marine ecosystems are complex scenes and hence difficult to tackle from a computer vision perspective. Automated technology to monitor the health of our oceans can facilitate in detecting and identifying marine species while freeing up experts from the repetitive task of manual annotation. Classification of coral species is a challenging task in itself and deep learning has a potential of solving this problem efficiently.

Programmable multi-task on-chip processing is proposed for improving the performance of CMOS imagers in terms of sensitivity adaptation and image processing capabilities. The current-mode fully analog on-chip processing only performs computations during the readout and analog-to-digital conversion phases, removing the need for any in-pixel or focal-plane processing circuitry. A VLSI implementation, in AMI 0.5 /spl mu/m CMOS process, results in significant silicon area savings as processing circuitry accounts for less than 20% of the imager prototype core area. Only three externally tunable parameters are required to fully define the processing task to be carried out by the 32/spl times/32 CMOS imager prototype, which performs sensitivity adaptation, edge detection or image enhancement on read-out.

We present a novel local surface description technique for automatic three dimensional (3D) object recognition. In the proposed approach, highly repeatable keypoints are first detected by computing the divergence of the vector field at each point of the surface. Being a differential invariant of curves and surfaces, the divergence captures significant information about the surface variations at each point. The detected keypoints are pruned to only retain the keypoints which are associated with high divergence values. A keypoint saliency measure is proposed to rank these keypoints and select the best ones. A novel integral invariant local surface descriptor, called 3D-Vor, is built around each keypoint by exploiting the vorticity of the vector field at each point of the local surface. The proposed descriptor combines the strengths of signature-based methods and integral invariants to provide robust local surface description. The performance of the proposed fully automatic 3D object recognition technique was rigorously tested on three publicly available datasets. Our proposed technique is shown to exhibit superior performance compared to state-of-the-art techniques. Our keypoint detector and descriptor based algorithm achieves recognition rates of 100%, 99.35% and 96.2% respectively, when tested on the Bologna, UWA and Ca׳ Foscari Venezia datasets.

We present a novel technique for image set based face/object recognition, where each gallery and query example contains a face/object image set captured from different viewpoints, background, facial expressions, resolution and illumination levels. While several image set classification approaches have been proposed in recent years, most of them represent each image set as a single linear subspace, mixture of linear subspaces or Lie group of Riemannian manifold. These techniques make prior assumptions in regards to the specific category of the geometric surface on which images of the set are believed to lie. This could result in a loss of discriminative information for classification. This paper alleviates these limitations by proposing an Iterative Deep Learning Model (IDLM) that automatically and hierarchically learns discriminative representations from raw face and object images. In the proposed approach, low level translationally invariant features are learnt by the Pooled Convolutional Layer (PCL). The latter is followed by Artificial Neural Networks (ANNs) applied iteratively in a hierarchical fashion to learn a discriminative non-linear feature representation of the input image sets. The proposed technique was extensively evaluated for the task of image set based face and object recognition on YouTube Celebrities, Honda/UCSD, CMU Mobo and ETH-80 (object) dataset, respectively. Experimental results and comparisons with state-of-the-art methods show that our technique achieves the best performance on all these datasets.

Interaction prediction has a wide range of applications such as robot controlling and prevention of dangerous events. In this paper, we introduce a new method to capture deep temporal information in videos for human interaction prediction. We propose to use flow coding images to represent the low-level motion information in videos and extract deep temporal features using a deep convolutional neural network architecture. We tested our method on the UT-Interaction dataset and the challenging TV human interaction dataset, and demonstrated the advantages of the proposed deep temporal features based on flow coding images. The proposed method, though using only the temporal information, outperforms the state of the art methods for human interaction prediction.

© 2016 IEEE. Healthy coral reefs play a vital role in maintaining biodiversity in tropical marine ecosystems. Deep sea exploration and imaging have provided us with a great opportunity to look into the vast and complex marine ecosystems. Data acquisition from the coral reefs has facilitated the scientific investigation of these intricate ecosystems. Millions of digital images of the sea floor have been collected with the help of Remotely Operated Vehicles (ROVs) and Autonomous Underwater Vehicles (AUVs). Automated technology to monitor the health of the oceans allows for transformational ecological outcomes by standardizing methods for detecting and identifying species. Manual annotation is a tediously repetitive and a time consuming task for marine experts. It takes 10-30 minutes for a marine expert to meticulously annotate a single image. This paper aims to automate the analysis of large available AUV imagery by developing advanced deep learning tools for rapid and large-scale automatic annotation of marine coral species. Such an automated technology would greatly benefit marine ecological studies in terms of cost, speed, accuracy and thus in better quantifying the level of environmental change marine ecosystems can tolerate. We propose a deep learning based classification method for coral reefs. We also report the application of the proposed technique towards the automatic annotation of unlabelled mosaics of the coral reef in the Abrolhos Islands, Western Australia. Our proposed method automatically quantifies the coral coverage in this region and detects a decreasing trend in coral population which is in line with conclusions by marine ecologists.

© 2016 IEEE. Coral reefs exhibit significant within-class variations, complex between-class boundaries and inconsistent image clarity. This makes coral classification a challenging task. In this paper, we report the application of generic CNN representations combined with hand-crafted features for coral reef classification to take advantage of the complementary strengths of these representation types. We extract CNN based features from patches centred at labelled pixels at multiple scales. We use texture and color based hand-crafted features extracted from the same patches to complement the CNN features. Our proposed method achieves a classification accuracy that is higher than the state-of-art methods on the MLC benchmark dataset for corals.

This paper tackles the problem of feature matching and range image registration. Our approach is based on a novel set of discriminating three-dimensional (3D) local features, named 3D-Vor (Vorticity). In contrast to conventional local feature representation techniques, which use the vector field (i.e. surface normals) to just construct their local reference frames, the proposed feature representation exploits the vorticity of the vector field computed at each point of the local surface to capture the distinctive characteristics at each point of the underlying 3D surface. The 3D-Vor descriptors of two range images are then matched using a fully automatic feature matching algorithm which identifies correspondences between the two range images. Correspondences are verified in a local validation step of the proposed algorithm and used for the pairwise registration of the range images. Quantitative results on low resolution Kinect 3D data (Washington RGB-D dataset) show that our proposed automatic registration algorithm is accurate and computationally efficient. The performance evaluation of the proposed descriptor was also carried out on the challenging low resolution Washington RGB-D (Kinect) object dataset, for the tasks of automatic range image registration. Reported experimental results show that the proposed local surface descriptor is robust to resolution, noise and more accurate than state-of-the-art techniques. It achieves 90% registration accuracy compared to 50%, 69.2% and 52% for spin image, 3D SURF and SISI/LD-SIFT descriptors, respectively.

In this paper we introduce sign constrained rectifier networks (SCRN), demonstrate their universal classification power and illustrate their applications to pattern decompositions. We prove that the proposed two-hidden-layer SCRN, with sign constraints on the weights of the output layer and on those of the top hidden layer, are capable of separating any two disjoint pattern sets. Furthermore, a two-hidden-layer SCRN of a pair of disjoint pattern sets can be used to decompose one of the pattern sets into several subsets so that each subset is convexly separable from the entire other pattern set; and a single-hidden-layer SCRN of a pair of convexly separable pattern sets can be used to decompose one of the pattern sets into several subsets so that each subset is linearly separable from the entire other pattern set. SCRN can thus be used to learn the pattern structures from the decomposed subsets of patterns and to analyse the discriminant factors of different patterns from the linear classifiers of the linearly separable subsets in the decompositions. With such pattern decompositions exhibiting convex separability or linear separability, users can also analyse the complexity of the classification problem, remove the outliers and the non-crucial points to improve the training of the traditional unconstrained rectifier networks in terms of both performance and efficiency.

To follow...

© Springer International Publishing Switzerland 2015. In this paper we introduce sign constrained rectifier networks (SCRN), demonstrate their universal classification power and illustrate their applications to pattern decompositions.We prove that the proposed two-hidden-layer SCRN, with sign constraints on the weights of the output layer and on those of the top hidden layer, are capable of separating any two disjoint pattern sets. Furthermore, a two-hidden-layer SCRN of a pair of disjoint pattern sets can be used to decompose one of the pattern sets into several subsets so that each subset is convexly separable from the entire other pattern set; and a single-hidden-layer SCRN of a pair of convexly separable pattern sets can be used to decompose one of the pattern sets into several subsets so that each subset is linearly separable from the entire other pattern set. SCRN can thus be used to learn the pattern structures from the decomposed subsets of patterns and to analyse the discriminant factors of different patterns from the linear classifiers of the linearly separable subsets in the decompositions. With such pattern decompositions exhibiting convex separability or linear separability, users can also analyse the complexity of the classification problem, remove the outliers and the non-crucial points to improve the training of the traditional unconstrained rectifier networks in terms of both performance and efficiency.

© 2015 IEEE. To find the optimal nonlinear separating boundary with maximum margin in the input data space, this paper proposes Contractive Rectifier Networks (CRNs), wherein the hidden-layer transformations are restricted to be contraction mappings. The contractive constraints ensure that the achieved separating margin in the input space is larger than or equal to the separating margin in the output layer. The training of the proposed CRNs is formulated as a linear support vector machine (SVM) in the output layer, combined with two or more contractive hidden layers. Effective algorithms have been proposed to address the optimization challenges arising from contraction constraints. Experimental results on MNIST, CIFAR-10, CIFAR-100 and MIT-67 datasets demonstrate that the proposed contractive rectifier networks consistently outperform their conventional unconstrained rectifier network counterparts.