Deep Convolutional Neural Networks (DCNNs) commonly use generic `max-pooling' (MP) layers to extract deformation-invariant features, but we argue in favor of a more refined treatment. First, we introduce epitomic convolution as a building block alternative to the common convolution-MP cascade of DCNNs; while having identical complexity to MP, Epitomic Convolution allows for parameter sharing across different filters, resulting in faster convergence and better generalization. Second, we introduce a Multiple Instance Learning approach to explicitly accommodate global translation and scaling when training a DCNN exclusively with class labels. For this we rely on a `patchwork' data structure that efficiently lays out all image scales and positions as candidates to a DCNN. Factoring global and local deformations allows a DCNN to `focus its resources' on the treatment of non-rigid deformations and yields a substantial classification accuracy improvement. Third, further pursuing this idea, we develop an efficient DCNN sliding window object detector that employs explicit search over position, scale, and aspect ratio. We provide competitive image classification and localization results on the ImageNet dataset and object detection results on the Pascal VOC 2007 benchmark.
This paper proposes to go beyond the state-of-the-art deep convolutional neural network (CNN) by incorporating the information from object detection, focusing on dealing with fine-grained image classification. Unfortunately, CNN suffers from over-fiting when it is trained on existing fine-grained image classification benchmarks, which typically only consist of less than a few tens of thousands training images. Therefore, we first construct a large-scale fine-grained car recognition dataset that consists of 333 car classes with more than 150 thousand training images. With this large-scale dataset, we are able to build a strong baseline for CNN with top-1 classification accuracy of 81.6%. One major challenge in fine-grained image classification is that many classes are very similar to each other while having large within-class variation. One contributing factor to the within-class variation is cluttered image background. However, the existing CNN training takes uniform window sampling over the image, acting as blind on the location of the object of interest. In contrast, this paper proposes an \emph{object-centric sampling} (OCS) scheme that samples image windows based on the object location information. The challenge in using the location information lies in how to design powerful object detector and how to handle the imperfectness of detection results. To that end, we design a saliency-aware object detection approach specific for the setting of fine-grained image classification, and the uncertainty of detection results are naturally handled in our OCS scheme. Our framework is demonstrated to be very effective, improving top-1 accuracy to 89.3% (from 81.6%) on the large-scale fine-grained car classification dataset.
Motivated by the remarkable successes of Graph-based Transduction (GT) and Sparse Representation (SR), we present a novel Classifier named Sparse Graph-based Classifier (SGC) for image classification. In SGC, SR is leveraged to measure the correlation (similarity) of each two samples and a graph is constructed for encoding these correlations. Then the Laplacian eigenmapping is adopted for deriving the graph Laplacian of the graph. Finally, SGC can be obtained by plugging the graph Laplacian into the conventional GT framework. In the image classification procedure, SGC utilizes the correlations, which are encoded in the learned graph Laplacian, to infer the labels of unlabeled images. SGC inherits the merits of both GT and SR. Compared to SR, SGC improves the robustness and the discriminating power of GT. Compared to GT, SGC sufficiently exploits the whole data. Therefore it alleviates the undercomplete dictionary issue suffered by SR. Four popular image databases are employed for evaluation. The results demonstrate that SGC can achieve a promising performance in comparison with the state-of-the-art classifiers, particularly in the small training sample size case and the noisy sample case.
Deep convolutional neural networks (CNN) has become the most promising method for object recognition, repeatedly demonstrating record breaking results for image classification and object detection in recent years. However, a very deep CNN generally involves many layers with millions of parameters, making the storage of the network model to be extremely large. This prohibits the usage of deep CNNs on resource limited hardware, especially cell phones or other embedded devices. In this paper, we tackle this model storage issue by investigating information theoretical vector quantization methods for compressing the parameters of CNNs. In particular, we have found in terms of compressing the most storage demanding dense connected layers, vector quantization methods have a clear gain over existing matrix factorization methods. Simply applying k-means clustering to the weights or conducting product quantization can lead to a very good balance between model size and recognition accuracy. For the 1000-category classification task in the ImageNet challenge, we are able to achieve 16-24 times compression of the network with only 1% loss of classification accuracy using the state-of-the-art CNN.
We present a technique for developing a network of re-used features, where the topology is formed using a coarse learning method, that allows gradient-descent fine tuning, known as an Abstract Deep Network (ADN). New features are built based on observed co-occurrences, and the network is maintained using a selection process related to evolutionary algorithms. This allows coarse ex- ploration of the problem space, effective for irregular domains, while gradient descent allows pre- cise solutions. Accuracy on standard UCI and Protein-Structure Prediction problems is comparable with benchmark SVM and optimized GBML approaches, and shows scalability for addressing large problems. The discrete implementation is symbolic, allowing interpretability, while the continuous method using fine-tuning shows improved accuracy. The binary multiplexer problem is explored, as an irregular domain that does not support gradient descent learning, showing solution to the bench- mark 135-bit problem. A convolutional implementation is demonstrated on image classification, showing an error-rate of 0.79% on the MNIST problem, without a pre-defined topology. The ADN system provides a method for developing a very sparse, deep feature topology, based on observed relationships between features, that is able to find solutions in irregular domains, and initialize a network prior to gradient descent learning.
We present highly efficient algorithms for performing forward and backward propagation of Convolutional Neural Network (CNN) for pixelwise classification on images. For pixelwise classification tasks, such as image segmentation and object detection, surrounding image patches are fed into CNN for predicting the classes of centered pixels via forward propagation and for updating CNN parameters via backward propagation. However, forward and backward propagation was originally designed for whole-image classification. Directly applying it to pixelwise classification in a patch-by-patch scanning manner is extremely inefficient, because surrounding patches of pixels have large overlaps, which lead to a lot of redundant computation. The proposed algorithms eliminate all the redundant computation in convolution and pooling on images by introducing novel d-regularly sparse kernels. It generates exactly the same results as those by patch-by-patch scanning. Convolution and pooling operations with such kernels are able to continuously access memory and can run efficiently on GPUs. A fraction of patches of interest can be chosen from each training image for backward propagation by applying a mask to the error map at the last CNN layer. Its computation complexity is constant with respect to the number of patches sampled from the image. Experiments have shown that our proposed algorithms speed up commonly used patch-by-patch scanning over 1500 times in both forward and backward propagation. The speedup increases with the sizes of images and patches.
In this paper, the problem of classifying synthetic and natural texture images is addressed. To tackle this problem, an innovative method is proposed that combines concepts from corrosion modeling and cellular automata to generate a texture descriptor. The core processes of metal (pitting) corrosion are identified and applied to texture images by incorporating the basic mechanisms of corrosion in the transition function of the cellular automaton. The surface morphology of the image is analyzed before and during the application of the transition function of the cellular automaton. In each iteration the cumulative mass of corroded product is obtained to construct each of the attributes of the texture descriptor. In a final step, this texture descriptor is used for image classification by applying Linear Discriminant Analysis. The method was tested on the well-known Brodatz and Vistex databases. In addition, in order to verify the robustness of the method, its invariance to noise and rotation were tested. To that end, different variants of the original two databases were obtained through addition of noise to and rotation of the images. The results showed that the method is effective for texture classification according to the high success rates obtained in all cases. This indicates the potential of employing methods inspired on natural phenomena in other fields.
The Convolutional Neural Network (CNN) has achieved great success in image classification. The classification model can also be utilized at image or patch level for many other applications, such as object detection and segmentation. In this paper, we propose a whole-image CNN regression model, by removing the full connection layer and training the network with continuous feature maps. This is a generic regression framework that fits many applications. We demonstrate this method through two tasks: simultaneous face detection & segmentation, and scene saliency prediction. The result is comparable with other models in the respective fields, using only a small scale network. Since the regression model is trained on corresponding image / feature map pairs, there are no requirements on uniform input size as opposed to the classification model. Our framework avoids classifier design, a process that may introduce too much manual intervention in model development. Yet, it is highly correlated to the classification network and offers some in-deep review of CNN structures.
Convolutional Neural Networks (CNNs) have achieved comparable error rates to well-trained human on ILSVRC2014 image classification task. To achieve better performance, the complexity of CNNs is continually increasing with deeper and bigger architectures. Though CNNs achieved promising external classification behavior, understanding of their internal work mechanism is still limited. In this work, we attempt to understand the internal work mechanism of CNNs by probing the internal representations in two comprehensive aspects, i.e., visualizing patches in the representation spaces constructed by different layers, and visualizing visual information kept in each layer. We further compare CNNs with different depths and show the advantages brought by deeper architecture.
Sparse coding can learn good robust representation to noise and model more higher-order representation for image classification. However, the inference algorithm is computationally expensive even though the supervised signals are used to learn compact and discriminative dictionaries in sparse coding techniques. Luckily, a simplified neural network module (SNNM) has been proposed to directly learn the discriminative dictionaries for avoiding the expensive inference. But the SNNM module ignores the sparse representations. Therefore, we propose a sparse SNNM module by adding the mixed-norm regularization (l1/l2 norm). The sparse SNNM modules are further stacked to build a sparse deep stacking network (S-DSN). In the experiments, we evaluate S-DSN with four databases, including Extended YaleB, AR, 15 scene and Caltech101. Experimental results show that our model outperforms related classification methods with only a linear classifier. It is worth noting that we reach 98.8% recognition accuracy on 15 scene.
This paper presents a new framework for visual bag-of-words (BOW) refinement and reduction to overcome the drawbacks associated with the visual BOW model which has been widely used for image classification. Although very influential in the literature, the traditional visual BOW model has two distinct drawbacks. Firstly, for efficiency purposes, the visual vocabulary is commonly constructed by directly clustering the low-level visual feature vectors extracted from local keypoints, without considering the high-level semantics of images. That is, the visual BOW model still suffers from the semantic gap, and thus may lead to significant performance degradation in more challenging tasks (e.g. social image classification). Secondly, typically thousands of visual words are generated to obtain better performance on a relatively large image dataset. Due to such large vocabulary size, the subsequent image classification may take sheer amount of time. To overcome the first drawback, we develop a graph-based method for visual BOW refinement by exploiting the tags (easy to access although noisy) of social images. More notably, for efficient image classification, we further reduce the refined visual BOW model to a much smaller size through semantic spectral clustering. Extensive experimental results show the promising performance of the proposed framework for visual BOW refinement and reduction.
Dynamic texture and scene classification are two fundamental problems in understanding natural video content. Extracting robust and effective features is a crucial step towards solving these problems. However the existing approaches suffer from the sensitivity to either varying illumination, or viewpoint changing, or even camera motion, and/or the lack of spatial information. Inspired by the success of deep structures in image classification, we attempt to leverage a deep structure to extract feature for dynamic texture and scene classification. To tackle with the challenges in training a deep structure, we propose to transfer some prior knowledge from image domain to video domain. To be specific, we propose to apply a well-trained Convolutional Neural Network (ConvNet) as a mid-level feature extractor to extract features from each frame, and then form a representation of a video by concatenating the first and the second order statistics over the mid-level features. We term this two-level feature extraction scheme as a Transferred ConvNet Feature (TCoF). Moreover we explore two different implementations of the TCoF scheme, i.e., the \textit{spatial} TCoF and the \textit{temporal} TCoF, in which the mean-removed frames and the difference between two adjacent frames are used as the inputs of the ConvNet, respectively. We evaluate systematically the proposed spatial TCoF and the temporal TCoF schemes on three benchmark data sets, including DynTex, YUPENN, and Maryland, and demonstrate that the proposed approach yields superior performance.
Promising results have been achieved in image classification problems by exploiting the discriminative power of sparse representations for classification (SRC). Recently, it has been shown that the use of \emph{class-specific} spike-and-slab priors in conjunction with the class-specific dictionaries from SRC is particularly effective in low training scenarios. As a logical extension, we build on this framework for multitask scenarios, wherein multiple representations of the same physical phenomena are available. We experimentally demonstrate the benefits of mining joint information from different camera views for multi-view face recognition.
This paper proposes a new method for vector quantization by minimizing the Kullback-Leibler Divergence between the class label distributions over the quantization inputs, which are original vectors, and the output, which is the quantization subsets of the vector set. In this way, the vector quantization output can keep as much information of the class label as possible. An objective function is constructed and we also developed an iterative algorithm to minimize it. The new method is evaluated on bag-of-features based image classification problem.
The ImageNet Large Scale Visual Recognition Challenge is a benchmark in object category classification and detection on hundreds of object categories and millions of images. The challenge has been run annually from 2010 to present, attracting participation from more than fifty institutions. This paper describes the creation of this benchmark dataset and the advances in object recognition that have been possible as a result. We discuss the challenges of collecting large-scale ground truth annotation, highlight key breakthroughs in categorical object recognition, provide a detailed analysis of the current state of the field of large-scale image classification and object detection, and compare the state-of-the-art computer vision accuracy with human accuracy. We conclude with lessons learned in the five years of the challenge, and propose future directions and improvements.
Rectified activation units (rectifiers) are essential for state-of-the-art neural networks. In this work, we study rectifier neural networks for image classification from two aspects. First, we propose a Parametric Rectified Linear Unit (PReLU) that generalizes the traditional rectified unit. PReLU improves model fitting with nearly zero extra computational cost and little overfitting risk. Second, we derive a robust initialization method that particularly considers the rectifier nonlinearities. This method enables us to train extremely deep rectified models directly from scratch and to investigate deeper or wider network architectures. Based on our PReLU networks (PReLU-nets), we achieve 4.94% top-5 test error on the ImageNet 2012 classification dataset. This is a 26% relative improvement over the ILSVRC 2014 winner (GoogLeNet, 6.66%). To our knowledge, our result is the first to surpass human-level performance (5.1%, Russakovsky et al.) on this visual recognition challenge.
In histopathological image analysis, feature extraction for classification is a challenging task due to the diversity of histology features suitable for each problem as well as presence of rich geometrical structure. In this paper, we propose an automatic feature discovery framework for extracting discriminative class-specific features and present a low-complexity method for classification and disease grading in histopathology. Essentially, our Discriminative Feature-oriented Dictionary Learning (DFDL) method learns class-specific features which are suitable for representing samples from the same class while are poorly capable of representing samples from other classes. Experiments on three challenging real-world image databases: 1) histopathological images of intraductal breast lesions, 2) mammalian lung images provided by the Animal Diagnostics Lab (ADL) at Pennsylvania State University, and 3) brain tumor images from The Cancer Genome Atlas (TCGA) database, show the significance of DFDL model in a variety problems over state-of-the-art methods
Sparse representation models a signal as a linear combination of a small number of dictionary atoms. As a generative model, it requires the dictionary to be highly redundant in order to ensure both a stable high sparsity level and a low reconstruction error for the signal. However, in practice, this requirement is usually impaired by the lack of labelled training samples. Fortunately, previous research has shown that the requirement for a redundant dictionary can be less rigorous if simultaneous sparse approximation is employed, which can be carried out by enforcing various structured sparsity constraints on the sparse codes of the neighboring pixels. In addition, numerous works have shown that applying a variety of dictionary learning methods for the sparse representation model can also improve the classification performance. In this paper, we highlight the task-driven dictionary learning algorithm, which is a general framework for the supervised dictionary learning method. We propose to enforce structured sparsity priors on the task-driven dictionary learning method in order to improve the performance of the hyperspectral classification. Our approach is able to benefit from both the advantages of the simultaneous sparse representation and those of the supervised dictionary learning. We enforce two different structured sparsity priors, the joint and Laplacian sparsity, on the task-driven dictionary learning method and provide the details of the corresponding optimization algorithms. Experiments on numerous popular hyperspectral images demonstrate that the classification performance of our approach is superior to sparse representation classifier with structured priors or the task-driven dictionary learning method.
Constructing effective representations is a critical but challenging problem in multimedia understanding. The traditional handcraft features often rely on domain knowledge, limiting the performances of exiting methods. This paper discusses a novel computational architecture for general image feature mining, which assembles the primitive filters (i.e. Gabor wavelets) into compositional features in a layer-wise manner. In each layer, we produce a number of base classifiers (i.e. regression stumps) associated with the generated features, and discover informative compositions by using the boosting algorithm. The output compositional features of each layer are treated as the base components to build up the next layer. Our framework is able to generate expressive image representations while inducing very discriminate functions for image classification. The experiments are conducted on several public datasets, and we demonstrate superior performances over state-of-the-art approaches.
Dictionary learning algorithms have been successfully used in both reconstructive and discriminative tasks, where the input signal is represented by a linear combination of a few dictionary atoms. While these methods are usually developed under $\ell_1$ sparsity constrain (prior) in the input domain, recent studies have demonstrated the advantages of sparse representation using structured sparsity priors in the kernel domain. In this paper, we propose a supervised dictionary learning algorithm in the kernel domain for hyperspectral image classification. In the proposed formulation, the dictionary and classifier are obtained jointly for optimal classification performance. The supervised formulation is task-driven and provides learned features from the hyperspectral data that are well suited for the classification task. Moreover, the proposed algorithm uses a joint ($\ell_{12}$) sparsity prior to enforce collaboration among the neighboring pixels. The simulation results illustrate the efficiency of the proposed dictionary learning algorithm.