Multilayer neural networks trained with the back-propagation algorithm constitute the best example of a successful gradient based learning technique. Given an appropriate network architecture, gradient-based learning algorithms can be used to synthesize a complex decision surface that can classify high-dimensional patterns, such as handwritten characters, with minimal preprocessing. This paper reviews various methods applied to handwritten character recognition and compares them on a standard handwritten digit recognition task. Convolutional neural networks, which are specifically designed to deal with the variability of 2D shapes, are shown to outperform all other techniques. Real-life document recognition systems are composed of multiple modules including field extraction, segmentation recognition, and language modeling. A new learning paradigm, called graph transformer networks (GTN), allows such multimodule systems to be trained globally using gradient-based methods so as to minimize an overall performance measure. Two systems for online handwriting recognition are described. Experiments demonstrate the advantage of global training, and the flexibility of graph transformer networks. A graph transformer network for reading a bank cheque is also described. It uses convolutional neural network character recognizers combined with global training techniques to provide record accuracy on business and personal cheques. It is deployed commercially and reads several million cheques per day.

Gradient-based learning applied to document recognition

We propose a deep convolutional neural network architecture codenamed Inception that achieves the new state of the art for classification and detection in the ImageNet Large-Scale Visual Recognition Challenge 2014 (ILSVRC14). The main hallmark of this architecture is the improved utilization of the computing resources inside the network. By a carefully crafted design, we increased the depth and width of the network while keeping the computational budget constant. To optimize quality, the architectural decisions were based on the Hebbian principle and the intuition of multi-scale processing. One particular incarnation used in our submission for ILSVRC14 is called GoogLeNet, a 22 layers deep network, the quality of which is assessed in the context of classification and detection.

/pdf/going-deeper-with-convolutions-1yobw2o2ds.pdf

Going deeper with convolutions

There is, I think, something ethereal about i —the square root of minus one. I remember first hearing about it at school. It seemed an odd beast at that time—an intruder hovering on the edge of reality.

Usually familiarity dulls this sense of the bizarre, but in the case of i it was the reverse: over the years the sense of its surreal nature intensified. It seemed that it was impossible to write mathematics that described the real world in …

I and i

We study the question of feature sets for robust visual object recognition; adopting linear SVM based human detection as a test case. After reviewing existing edge and gradient based descriptors, we show experimentally that grids of histograms of oriented gradient (HOG) descriptors significantly outperform existing feature sets for human detection. We study the influence of each stage of the computation on performance, concluding that fine-scale gradients, fine orientation binning, relatively coarse spatial binning, and high-quality local contrast normalization in overlapping descriptor blocks are all important for good results. The new approach gives near-perfect separation on the original MIT pedestrian database, so we introduce a more challenging dataset containing over 1800 annotated human images with a large range of pose variations and backgrounds.

/pdf/histograms-of-oriented-gradients-for-human-detection-3nteinziuq.pdf

Histograms of oriented gradients for human detection

Object detection performance, as measured on the canonical PASCAL VOC dataset, has plateaued in the last few years. The best-performing methods are complex ensemble systems that typically combine multiple low-level image features with high-level context. In this paper, we propose a simple and scalable detection algorithm that improves mean average precision (mAP) by more than 30% relative to the previous best result on VOC 2012 -- achieving a mAP of 53.3%. Our approach combines two key insights: (1) one can apply high-capacity convolutional neural networks (CNNs) to bottom-up region proposals in order to localize and segment objects and (2) when labeled training data is scarce, supervised pre-training for an auxiliary task, followed by domain-specific fine-tuning, yields a significant performance boost. Since we combine region proposals with CNNs, we call our method R-CNN: Regions with CNN features. We also present experiments that provide insight into what the network learns, revealing a rich hierarchy of image features. Source code for the complete system is available at http://www.cs.berkeley.edu/~rbg/rcnn.

/pdf/rich-feature-hierarchies-for-accurate-object-detection-and-22fvjlggyg.pdf

Rich Feature Hierarchies for Accurate Object Detection and Semantic Segmentation

Previous research in automatic facial expression recognition has been limited to recognition of gross expression categories (e.g., joy or anger) in posed facial behavior under well-controlled conditions (e.g., frontal pose and minimal out-of-plane head motion). We have developed a system that detects a discrete and important facial action (e.g., eye blinking) in spontaneously occurring facial behavior that has been measured with a nonfrontal pose, moderate out-of-plane head motion, and occlusion. The system recovers three-dimensional motion parameters, stabilizes facial regions, extracts motion and appearance information, and recognizes discrete facial actions in spontaneous facial behavior. We tested the system in video data from a two-person interview. The 10 subjects were ethnically diverse, action units occurred during speech, and out-of-plane motion and occlusion from head motion and glasses were common. The video data were originally collected to answer substantive questions in psychology and represent a substantial challenge to automated action unit recognition. In analysis of blinks, the system achieved 98% accuracy.

/pdf/automatic-recognition-of-eye-blinking-in-spontaneously-2hu04ieyr0.pdf

Automatic recognition of eye blinking in spontaneously occurring behavior.

In this paper we demonstrate two new approaches to deriving three-dimensional surface orientation information (“shape”) from two-dimensional image cues. The two approaches are the method of affine-transformable patterns and the shape-from-texture paradigm. They are introduced by a specific application common to both: the concept of skewed symmetry. Skewed symmetry is shown to constrain the relationship of observed distortions in a known object regularity to a small subset of possible underlying surface orientations. Besides this constraint, valuable in its own right, the two methods are shown to generate other surface constraints as well. Some applications are presented of skewed symmetry to line drawing analysis, to the use of gravity in shape understanding, and to global shape recovery.

/pdf/mapping-image-properties-into-shape-constraints-skewed-2qx80fwcyh.pdf

Mapping image properties into shape constraints: skewed symmetry, affine-transformable patterns, and the shape-from-texture paradigm

Detecting and segmenting cell regions in microscopic images is a challenging task, because cells typically do not have rich features, and their shapes and appearances are highly irregular and flexible. Furthermore, cells often form clusters, rendering the existing joint detection and segmentation algorithms unable to segment out individual cells. We address these difficulties by proposing a Heterogeneous Conditional Random Field (HCRF), in which different nodes have different state sets. The state sets are designed in such a way that the resulting HCRF model could encode all possible detection/segmentation cases while keeping its identifiability and compactness. Attributed to the provably optimal design of the state sets, the proposed model successfully realizes joint detection and segmentation of the cell regions into individual cells whether the cells are separate or touch one another. Experiments on two different types of cell images show that the HCRF outperforms several recently proposed methods.

/pdf/heterogeneous-conditional-random-field-realizing-joint-ycsig9efz3.pdf

Heterogeneous Conditional Random Field: Realizing joint detection and segmentation of cell regions in microscopic images

This paper describes a calibration method for determining the physical location of the ultrasound (US) image plane relative to a rigidly attached 3D position sensor. A calibrated US probe can measure the 3D spatial location of anatomic structures relative to a global coordinate system. The calibration is performed by aiming the US probe at a calibration target containing a known point (1 mm diameter sphere) in physical space. This point is repeatedly collected at various locations in the US image plane to produce the calibration dataset. An idealized model of the collection process is used to eliminate outliers from the calibration dataset and also to examine the theoretical accuracy limits of this method. The results demonstrate accurate and robust calibration of the 3D spatial relationship between the US image plane and the 3D position sensor.

/pdf/calibration-method-for-determining-the-physical-location-of-4o8ge5dzlo.pdf

Calibration Method for Determining the Physical Location of the Ultrasound Image Plane

Measurement of human body is useful for ergonomic design in manufacturing. We aim to accurately measure the shapes of human feet for the design of shoes, for which measuring the dynamic shape of the foot in motion is important because the foot deforms while walking or running. In this paper, we propose a method for measuring the anatomical feature cross-sections of the foot while walking. The dynamic shape of feature cross-section has never been measured, though those features are very basic and popular. Our proposed method is based on the multi-view stereo method. The target cross-sections are painted in individual colors (red, green, yellow and blue). Several nonlinear conditions are introduced in the process to find the consistent correspondence in all images. We show that the proposed method for dynamic measurement achieved the desired accuracy similar to existing 3D static foot scanners.

Takeo Kanade

Papers

Automatic recognition of eye blinking in spontaneously occurring behavior.

Mapping image properties into shape constraints: skewed symmetry, affine-transformable patterns, and the shape-from-texture paradigm

Heterogeneous Conditional Random Field: Realizing joint detection and segmentation of cell regions in microscopic images

Calibration Method for Determining the Physical Location of the Ultrasound Image Plane

3D measurement of feature cross-sections of foot while walking