Integrated Segmentation and Recognition of Hand-Printed Numerals

By J. Keeler, David E. Rumelhart, Wee Kheng Leow. 1990

Subfields: Computer Vision and Pattern Recognition

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Computer Vision and Pattern Recognition

Citations in the corpus

Unified Perceptual Parsing for Scene Understanding (2018, background): “For example, a number of previous research works focus on the combination of recognition and segmentation [19,20,21].”
UberNet: Training a Universal Convolutional Neural Network for Low-, Mid-, and High-Level Vision Using Diverse Datasets and Limited Memory (2016, background): “Finally, the particular motivation for this research has been the interest in studying the synergy between different visual tasks (e.g. the long-standing problem of combining segmentation and recognition [9, 47, 50, 53, 68, 72, 99]), so this work can be understood as a ﬁrst step in this direction.”
Attention-based Deep Multiple Instance Learning (2018, neutral): “As shown in Theorem 1 and 2, there are two MIL pooling operators that ensure the score function ( i.e. , the bag probability) to be a symmetric function, namely, the maximum operator: and the mean operator: In fact, other operators could be used such as, the convex maximum operator ( i.e. , log-sum-”
Multi-Instance Kernels (2002, methodology): “However, similar problems and algorithms have been considered earlier, for exam- ple in pattern recognition (Keeler et al., 1991).”
Deep Multi-patch Aggregation Network for Image Style, Aesthetics, and Quality Estimation (2015, methodology): “Recent studies have attempted to unify deep learning and the MIL framework for many applications.”
Is object localization for free? - Weakly-supervised learning with convolutional neural networks (2015, unknown): “Second, we introduce a max-pooling layer that hypothesizes the possible location of the object in the image, similar to [32, Section 4] and [28].”
Competition and Multiple Cause Models (1995, background): “and substituting these into equation 3, we get pCj = Xi sicijQl6=i(1 slclj) Ql(1 slclj) +Pk skckjQl6=k(1 slclj) (6) = 1 1 1+Xk skckj 1 ckj (7) using the facts that the ratio of equations 4 and 5 is just the odds sicij 1 cij that cause i generates j, and si is either 0 or 1.”
Using Generative Models for Handwritten Digit Recognition (1996, background): “More recently there have been a number [7], [8], [9] of successful attempts to automatically learn appropriate lo-cal features using feed forward neural networks.”
Model-Based Multiple Instance Learning (2017, background): “Stemming from research on handwritten digit recognition in 1990 [3], MI learning has emerged as an important topic in pattern recognition with a host of techniques and applications, see e.”
Multiple Instance Boosting for Object Detection (2005, methodology): “The idea for multiple instance learning was originally proposed 1990 for handwritten digit recognition by Keeler, et. al [7].”
Gradient-based learning applied to document recognition (1998, methodology): “If the partial derivative of E p with respect to Xn is known, then the partial derivatives of E p with respect to Wn and Xn(cid:0)1 can be computed using the backward recurrence @E p @Wn = @F @W (Wn;Xn(cid:0)1) @E p @Xn @E p @Xn(cid:0)1 = @F”
Estimating a Kernel Fisher Discriminant in the Presence of Label Noise (2001, methodology): “For an alternative approach to this type of problem see (Keeler et al., 1991; Maron & Lozano-Prez, 1998).”
Varieties of Helmholtz Machine (1996, methodology): “Indeed, in the deterministic version of the machine (Dayan et al, 1995) that we used for solving the 8-bit shifter problem, we used a particular competitive imaging model (Saund, 1995; Dayan & Zemel, 1995) for the output units that was inspired by the Integrated Segmentation and Recognition architec”
Handling label noise in video classification via multiple instance learning (2011, background): “Other models, such as average probability, or the ISR criterion in [17], might yield better performance.”