General Recognition Problem Research Articles

A CNN-ELM Classification Model for Automated Tomato Maturity Grading

Tomatoes are popular around the world due to their high nutritional value. Tomatoes are also one of the world’s most widely cultivated and profitable crops. The distribution and marketing of tomatoes depend highly on their quality. Estimating tomato ripeness is an essential step in determining shelf life and quality. With the abundant supply of tomatoes on the market, it is exceedingly difficult to estimate tomato ripeness using human graders. To address this issue and improve tomato quality inspection and sorting, automated tomato maturity classification models based on different features have been developed. However, current methods heavily rely on human-engineered or handcrafted features. Convolutional neural networks have emerged as the preferred technique for general object recognition problems because they can automatically detect and extract valuable features by directly working on input images. This paper proposes a CNN-ELM classification model for automated tomato maturity grading that combines CNNs’ automated feature learning capabilities with the efficiency of extreme learning machines to perform fast and accurate classification even with limited training data. The results showed that the proposed CNN-ELM model had a classification accuracy of 96.67% and an F1-score of 96.67% in identifying six maturity stages from the test data.

Separable discrete functions: Recognition and sufficient conditions

A discrete function of n variables is a mapping g:X1×…×Xn→A, where X1,…,Xn, and A are arbitrary finite sets. Function g is called separable if there exist n functions gi:Xi→A for i=1,…,n, such that for every input x1,…,xn the function g(x1,…,xn) takes one of the values g1(x1),…,gn(xn). Given a discrete function g, it is an interesting problem to ask whether g is separable or not. Although this seems to be a very basic problem concerning discrete functions, the complexity of recognition of separable discrete functions of n variables is known only for n=2. In this paper we will show that a slightly more general recognition problem, when g is not fully but only partially defined, is NP-complete for n≥3. We will then use this result to show that the recognition of fully defined separable discrete functions is NP-complete for n≥4.The general recognition problem contains the above mentioned special case for n=2. This case is well-studied in the context of game theory, where (separable) discrete functions of n variables are referred to as (assignable) n-person game forms. There is a known sufficient condition for assignability (separability) of two-person game forms (discrete functions of two variables) called (weak) total tightness of a game form. This property can be tested in polynomial time, and can be easily generalized both to higher dimension and to partially defined functions. We will prove in this paper that weak total tightness implies separability for (partially defined) discrete functions of n variables for any n, thus generalizing the above result known for n=2. Our proof is constructive. Using a graph-based discrete algorithm we show how for a given weakly totally tight (partially defined) discrete function g of n variables one can construct separating functions g1,…,gn in polynomial time with respect to the size of the input function.

NIC-planar graphs

A graph is NIC-planar if it admits a drawing in the plane with at most one crossing per edge and such that two pairs of crossing edges share at most one common end vertex. NIC-planarity generalizes IC-planarity, which allows a vertex to be incident to at most one crossing edge, and specializes 1-planarity, which only requires at most one crossing per edge.We characterize embeddings of maximal NIC-planar graphs in terms of generalized planar dual graphs. The characterization is used to derive tight bounds on the density of maximal NIC-planar graphs which ranges between 3.2(n−2) and 3.6(n−2). Further, we prove that optimal NIC-planar graphs with 3.6(n−2) edges have a unique embedding and can be recognized in linear time, whereas the general recognition problem of NIC-planar graphs is NP-complete. In addition, we show that there are NIC-planar graphs that do not admit right angle crossing drawings, which distinguishes NIC-planar from IC-planar graphs.

Odor Recognition in Multiple E-Nose Systems With Cross-Domain Discriminative Subspace Learning

In this paper, we propose an odor recognition framework for multiple electronic noses (E-noses), machine olfaction odor perception systems. Straight to the point, the proposed transferring odor recognition model is called cross-domain discriminative subspace learning (CDSL). General odor recognition problems with E-nose are single domain oriented, that is, recognition algorithms are often modeled and tested on the same one domain data set (i.e., from only one E-nose system). Different from that, we focus on a more realistic scenario: the recognition model is trained on a prepared source domain data set from a master E-nose system A, but tested on another target domain data set from a slave system B or C with the same type of the master system A. The internal device parameter variance between master and slave systems often results in data distribution discrepancy between source domain and target domain, such that single-domain-based odor recognition model may not be adapted to another domain. Therefore, we propose domain-adaptation-based odor recognition for addressing the realistic recognition scenario across systems. Specifically, the proposed CDSL method consists of three merits: 1) an intraclass scatter minimizationand an interclass scatter maximization-based discriminative subspace learning is solved on source domain; 2) a data fidelity and preservation constraint of the subspace is imposed on target domain without distortion; and 3) a minipatch feature weighted domain distance is minimized for closely connecting the source and target domains. Experiments and comparisons on odor recognition tasks in multiple E-noses demonstrate the efficiency of the proposed method.

Multi-Instance Deep Learning: Discover Discriminative Local Anatomies for Bodypart Recognition.

In general image recognition problems, discriminative information often lies in local image patches. For example, most human identity information exists in the image patches containing human faces. The same situation stays in medical images as well. "Bodypart identity" of a transversal slice-which bodypart the slice comes from-is often indicated by local image information, e.g., a cardiac slice and an aorta arch slice are only differentiated by the mediastinum region. In this work, we design a multi-stage deep learning framework for image classification and apply it on bodypart recognition. Specifically, the proposed framework aims at: 1) discover the local regions that are discriminative and non-informative to the image classification problem, and 2) learn a image-level classifier based on these local regions. We achieve these two tasks by the two stages of learning scheme, respectively. In the pre-train stage, a convolutional neural network (CNN) is learned in a multi-instance learning fashion to extract the most discriminative and and non-informative local patches from the training slices. In the boosting stage, the pre-learned CNN is further boosted by these local patches for image classification. The CNN learned by exploiting the discriminative local appearances becomes more accurate than those learned from global image context. The key hallmark of our method is that it automatically discovers the discriminative and non-informative local patches through multi-instance deep learning. Thus, no manual annotation is required. Our method is validated on a synthetic dataset and a large scale CT dataset. It achieves better performances than state-of-the-art approaches, including the standard deep CNN.

Good recognition is non-metric

Recognition is the fundamental task of visual cognition, yet how to formalize the general recognition problem for computer vision remains an open issue. The problem is sometimes reduced to the simplest case of recognizing matching pairs, often structured to allow for metric constraints. However, visual recognition is broader than just pair-matching: what we learn and how we learn it has important implications for effective algorithms. In this review paper, we reconsider the assumption of recognition as a pair-matching test, and introduce a new formal definition that captures the broader context of the problem. Through a meta-analysis and an experimental assessment of the top algorithms on popular data sets, we gain a sense of how often metric properties are violated by recognition algorithms. By studying these violations, useful insights come to light: we make the case for local distances and systems that leverage outside information to solve the general recognition problem.

Recognizing Plans with Loops Represented in a Lexicalized Grammar

This paper extends existing plan recognition research to handle plans containing loops. We supply an encoding of plans with loops for recognition, based on techniques used to parse lexicalized grammars, and demonstrate its effectiveness empirically. To do this, the paper first shows how encoding plan libraries as context free grammars permits the application of standard rewriting techniques to remove left recursion and ε-productions, thereby enabling polynomial time parsing. However, these techniques alone fail to provide efficient algorithms for plan recognition. We show how the loop-handling methods from formal grammars can be extended to the more general plan recognition problem and provide a method for encoding loops in an existing plan recognition system that scales linearly in the number of loop iterations.

Specificity of impaired facial identity recognition in children with suspected developmental prosopagnosia

Adults experiencing face recognition difficulties in the absence of known brain injury are described as cases of developmental prosopagnosia (DP), under the assumption that specific face recognition impairments have always been present. However, only five childhood cases of DP have been reported, and the majority had additional socio-communicative impairments consistent with an autism spectrum disorder (ASD). We tested face recognition skills of six 4- to 8-year-old children, who were suspected of having DP, and tested for ASD using established diagnostic tools. Two children met criteria for ASD. One child did not exhibit consistent face recognition impairments. The remaining three children were severely impaired on multiple tasks of unfamiliar face recognition despite normal cognitive functioning and no evidence of ASD. Two of these children were also impaired at object recognition suggesting more general visual recognition problems. The final child showed normal object recognition demonstrating apparently specific problems with facial identity recognition.

A correspondence-based neural mechanism for position invariant feature processing

We here focus on constructing a hierarchical neural sys-tem for position-invariant recognition, which is one of themost fundamental invariant recognition achieved in vis-ual processing [1,2]. The invariant recognition have beenhypothesized to be done by matching a sensory image ofa particular object stimulated on the retina to the mostsuitable representation stored in memory of the highervisual cortical area. Here arises a general problem: In sucha visual processing, the position of the object image onthe retina must be initially uncertain. Furthermore, theretinal activities possessing sensory information are beingfar from the ones in the higher area with a loss of the sen-sory object information. Nevertheless, with such recogni-tion ambiguity, the particular object can effortlessly andeasily be recognized. Our aim in this work is an attempt toresolve such a general recognition problem.

Visual homing in the absence of feature-based landmark information

Despite that fact that landmarks play a prominent role in human navigation, experimental evidence on how landmarks are selected and defined by human navigators remains elusive. Indeed, the concept of a ‘landmark’ is itself not entirely clear. In everyday language, the term landmark refers to salient, distinguishable, and usually nameable objects, rendering the problem of landmark recognition a special case of the general object recognition problem. In contrast, in the insect and robot literature, this notion of landmarks is often replaced by the “local position information” [e.g., Trullier, O., Wiener, S., Berthoz, A., & Meyer, J.-A. (1997). Biologically based artificial navigation systems: Review and prospects. Progress in Neurobiology, 51, 483–544], referring to the entire set of sensor readings obtained at one location. This set may then serve as a characteristic of the particular location. Honey bees have been shown to base place recognition and homing on a snapshot-like memory of the place’s visual environment, not on the distances to recognized objects [Cartwright, B., & Collett, T. (1983). Landmark learning in bees. Experiments and models. Journal of Comparative Physiology A, 151, 521–543]. A number of theoretical models of snapshot-based homing [e.g., Franz, M., Schölkopf, B., Mallot, H. A., Bülthoff, H. H. (1998). Where did I take that snapshot? Scene-based homing by image matching. Biological Cybernetics, 79, 191–202; Vardy, A., & Möller, R. (2005). Biologically plausible visual homing methods based on optical flow techniques. Connection Science, 17, 47–89] predict that the accuracy of snapshot-based homing should depend on image contrast. For rodent hippocampal place fields, models have been developed using additional image information such as three-dimensional depth and allocentric orientations (e.g., room axes) and are thus less sensitive to image contrast and noise [e.g. Barry, C., Lever, C., Hayman, R., Hartley, T., Burton, S., O’Keefe, J., et al. (2006). The boundary vector cell model of place cell firing and spatial memory. Reviews in the Neurosciences, 17, 71–79]. Here, we study human visual homing in a virtual environment void of objects and readily detected image features. The environment was a circular room with a homogenous colour gradient covering the wall and uniform floor and ceiling. Subjects were able to approach remembered places. Accuracy decreased with colour gradient modulation and room size, in qualitative agreement with the snapshot model but not with other models of place recognition. We conclude that human memory for places can make use of a snapshot algorithm.

A characterization of cycle-free unit probe interval graphs

A graph is a probe interval graph (PIG) if its vertices can be partitioned into probes and nonprobes with an interval assigned to each vertex so that vertices are adjacent if and only if their corresponding intervals overlap and at least one of them is a probe. PIGs are a generalization of interval graphs introduced by Zhang for an application concerning the physical mapping of DNA in the human genome project. PIGs have been characterized in the cycle-free case by Sheng, and other miscellaneous results are given by McMorris, Wang, and Zhang. Johnson and Spinrad give a polynomial time recognition algorithm for when the partition of vertices into probes and nonprobes is given. The complexity for the general recognition problem is not known. Here, we restrict attention to the case where all intervals have the same length, that is, we study the unit probe interval graphs and characterize the cycle-free graphs that are unit probe interval graphs via a list of forbidden induced subgraphs.

Networks with small stretch number

In a previous work, the authors introduced the class of graphs with bounded induced distance of order k (BID( k) for short), to model non-reliable interconnection networks. A network modeled as a graph in BID( k) can be characterized as follows: if some nodes have failed, as long as two nodes remain connected, the distance between these nodes in the faulty graph is at most k times the distance in the non-faulty graph. The smallest k such that G∈BID( k) is called stretch number of G. We show an odd characteristic of the stretch numbers: every rational number greater or equal 2 is a stretch number, but only discrete values are admissible for smaller stretch numbers. Moreover, we give a new characterization of classes BID(2−1/ i), i⩾1, based on forbidden induced subgraphs. By using this characterization, we provide a polynomial time recognition algorithm for graphs belonging to these classes, while the general recognition problem is Co-NP-complete.

Hypercolumn model: A combination model of hierarchical self-organizing maps and neocognitron for image recognition

In this paper, a neural network model, the hypercolumn model (HCM), which is applicable to general image recognition, is proposed. The HCM is a combination model of hierarchical self-organizing maps (HSOM) and neocognitron (NC); it resolves the disadvantages of both the HSOM and the NC, and inherits all of the advantages of both models. The HSOM quantizes and nonlinearly maps an input space into a small dimensional feature map. Using this feature map, the HSOM can function as an unsupervised clustering method and can easily solve pattern recognition problems, which are hard for traditional linear methods or for multilayered perceptrons using the back-propagation learning method because of the complex shapes of the class boundaries. On the other hand, the NC is a multilayered network, in which layers to extract local features and layers to integrate shifted features alternate. Using this structure, the NC can reduce the dimensionality gradually and recognize shifted and scaled patterns. By combining the characteristics of the HSOM and the structure of the NC, the HCM can be made applicable to general image recognition problems which have high dimensional input data and in which the shapes of the class boundaries are quite complex because of the high abstraction level of the classification. © 2000 Scripta Technica, Syst Comp Jpn, 31(2): 49–61, 2000

Improving discrimination of confusable words using the divergence measure

This paper describes the use of the divergence measure as a criterion for finding a transformation matrix which will map the original speech observations onto a subspace with more discriminative ability than the original. A gradient-based algorithm is also proposed to compute the transformation matrix efficiently. The subspace approach is used as a preprocessing step in a hidden Markov model (HMM) based system to enhance discrimination of acoustically similar pairs of words. This approach is compared with standard linear discriminant analysis (LDA) techniques and shown to yield as much as 4.5% improvement. The subspace approach is also applied successfully to a more general recognition problem, i.e., discrimination of K confusable words, using the average divergence measure.

Improving discrimination of confusable words using the divergence measure

This paper describes the use of the divergence measure as a criterion for finding a transformation matrix which will map the original speech observations onto a subspace with more discriminative ability than the original. A gradient-based algorithm is also proposed to compute the transformation matrix efficiently. The subspace approach is used as a preprocessing step in a hidden Markov model (HMM) based system to enhance discrimination of acoustically similar pairs of words. This approach is compared with standard linear discriminant analysis (LDA) techniques and shown to yield as much as 4.5% improvement. The subspace approach is also applied successfully to a more general recognition problem, i.e., discrimination of K confusable words, using the average divergence measure.

A Characterization of Nonnegative Box-Greedy Matrices

Given an ordering of the variables according to nonincreasing coefficients of the objective function $c^T x$, the nonnegative matrix A is said to be greedy if, under arbitrary nonnegative constraint vectors b and h, the greedy algorithm maximizes $c^T x$ subject to $Ax \leq b,0 \leq x \leq h$. Extending a result of Hoffman, Kolen, and Sakarovitch for $(0,1)$-matrices, we characterize greedy matrices in terms of forbidden submatrices, which yields polynomial recognition algorithms for various classes of greedy matrices. The general recognition problem for the existence of forbidden submatrices is shown to be NP-complete.

From HMM's to segment models: a unified view of stochastic modeling for speech recognition

Many alternative models have been proposed to address some of the shortcomings of the hidden Markov model (HMM), which is currently the most popular approach to speech recognition. In particular, a variety of models that could be broadly classified as segment models have been described for representing a variable-length sequence of observation vectors in speech recognition applications. Since there are many aspects in common between these approaches, including the general recognition and training problems, it is useful to consider them in a unified framework. The paper describes a general stochastic model that encompasses most of the models proposed in the literature, pointing out similarities of the models in terms of correlation and parameter tying assumptions, and drawing analogies between segment models and HMMs. In addition, we summarize experimental results assessing different modeling assumptions and point out remaining open questions.

On the Recognition of Information With a Digital Computer

There exists a vast discrepancy in the power of discrimination exercised by a digital computer and in that of a human being. The recognition of information or data patterns is a simple task for the least experienced human clerk. Most people possess sufficiently sophisticated recognition capabilities so that a variation in the pattern of the data under scrutiny will not cause undue difficulty in the discrimination process. The recognition powers of a digital computer, however, are best demonstrated in an elementary table lookup operation, wherein the subject information is required to match exactly with a portion of the master list in order to be “recognized”. Machine recognition of data which is allowed to vary from the predetermined digital pattern is a vastly more complex problem. Since digital computers are inherently devices which are capable of only YES or NO answers, all MAYBE or PERHAPS responses are obtained only through painstaking effort. If the variations in the subject data are allowed a reasonable range in both position and type, there is no complete solution of the recognition problem available with present techniques and equipments. This paper will outline the general recognition problem in terms of a set of definitions and a mathematical model. I believe that a useful formulation of the problem, and a comprehension of the difficulties involved in discrimination, are prerequisites to any effort at obtaining a complete solution to the problem of data recognition.