US20130260345A1 - Food recognition using visual analysis and speech recognition - Google Patents
Food recognition using visual analysis and speech recognition Download PDFInfo
- Publication number
- US20130260345A1 US20130260345A1 US13/849,222 US201313849222A US2013260345A1 US 20130260345 A1 US20130260345 A1 US 20130260345A1 US 201313849222 A US201313849222 A US 201313849222A US 2013260345 A1 US2013260345 A1 US 2013260345A1
- Authority
- US
- United States
- Prior art keywords
- food
- images
- color
- feature
- applying
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Abandoned
Links
Images
Classifications
-
- G—PHYSICS
- G09—EDUCATION; CRYPTOGRAPHY; DISPLAY; ADVERTISING; SEALS
- G09B—EDUCATIONAL OR DEMONSTRATION APPLIANCES; APPLIANCES FOR TEACHING, OR COMMUNICATING WITH, THE BLIND, DEAF OR MUTE; MODELS; PLANETARIA; GLOBES; MAPS; DIAGRAMS
- G09B19/00—Teaching not covered by other main groups of this subclass
- G09B19/0092—Nutrition
Landscapes
- Business, Economics & Management (AREA)
- Engineering & Computer Science (AREA)
- Health & Medical Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Nutrition Science (AREA)
- Entrepreneurship & Innovation (AREA)
- Physics & Mathematics (AREA)
- Educational Administration (AREA)
- Educational Technology (AREA)
- General Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Image Analysis (AREA)
Abstract
A method and system for analyzing at least one food item on a food plate is disclosed. A plurality of images of the food plate is received by an image capturing device. A description of the at least one food item on the food plate is received by a recognition device. The description is at least one of a voice description and a text description. At least one processor extracts a list of food items from the description; classifies and segments the at least one food item from the list using color and texture features derived from the plurality of images; and estimates the volume of the classified and segmented at least one food item. The processor is also configured to estimate the caloric content of the at least one food item.
Description
- This application is a continuation of pending U.S. patent application Ser. No. 12/683,124, filed on Jan. 6, 2010 which claims the benefit of U.S. provisional patent application No. 61/143,081 filed Jan. 7, 2009, the disclosure of which is incorporated herein by reference in its entirety.
- This invention was made with U.S. government support under contract number NIH 1U01HL091738-01. The U.S. government has certain rights in this invention.
- The invention relates generally to vision systems. More specifically, the invention relates to a system and method for automatically identifying items of food on a plate and computing the volume of each food item to aid in making a determination of the caloric content of the food on the plate.
- Studies have shown that a healthy diet can significantly reduce the risk of disease. This may provide a motivation, either self-initiated or from a doctor, to monitor and assess dietary intake in a systematic way. It is known that individuals do a poor job of assessing their true dietary intake. In the kitchen when preparing a meal, one can estimate the total caloric content of a meal by looking at food labels and calculating portion size, given a recipe of amounts of ingredients. At a restaurant, estimating caloric content of a meal is more difficult. A few restaurants may list in their menus the calorie value of certain low fat/dietary conscience meals, but the majority of meals are much higher in calories, so they are not listed. Even dieticians need to perform complex lab measurements to accurately assess caloric content of foods.
- Human beings are good at identifying food, such as the individual ingredients of a meal, but are known to be poor at volume estimation, and it is nearly impossible even of one had the total volume of a meal to estimate the volume of individual ingredients, which may be mixed and either seen or unseen. It is difficult for an individual to measure nutritional consumption by individuals in an easy yet quantitative manner. Several software applications, such as CalorieKing™, CaloricCounter™, etc., are of limited value since they perform a simple calculation based on portion size which cannot be accurately estimated by users. Veggie Vision™ claims to automatically recognize fruits and vegetables in a supermarket environment during food checkout. However, there are few, if any, published technical details about how this is achieved.
- Automatic image analysis techniques of the prior art are more successful at volume computation than at food item identification. Automated and accurate food recognition is particularly challenging because there are a large number of food types that people consume. A single category of food may have large variations. Moreover, diverse lighting conditions may greatly alter the appearance of food to a camera which is configured to a capture food appearance data. In F. Zhu et al., “Technology-assisted dietary assessment,” SPIE, 2008, (“hereinafter “Zhu et al.”), Zhu et al. uses an intensity-based segmentation and classification of each food item using color and texture features. Unfortunately, the system of Zhu et al. does not estimate the volume of food needed for accurate assessment of caloric content. State of the art object recognition methods, such as the methods described in M. Everingham et al., “The PASCAL Visual Object Classes Challenge 2008 (VOC2008),” are unable to operate on a large number of food classes.
- Recent success in recognition is largely due to the use of powerful image features and their combinations. Concatenated feature vectors are commonly used as input for classifiers. Unfortunately, this is feasible only when the features are homogeneous, e.g., as in the concatenation of two histograms (HOG and IMH) in N. Dalal et al., “Human detection using oriented histograms of flow and appearance,” ECCV, 2008. Linear combinations of multiple non-linear kernels, each of which is based on one feature type, is a more general way to integrate heterogeneous features, as in M. Varna and D. Ray, “Learning the discriminative power invariance tradeoff,” ICCV, 2007. However, both the vector concatenation and the kernel combination based methods require computation of all of the features.
- Accordingly, what would be desirable, but has not yet been provided, is a system and method for effective and automatic food recognition for large numbers of food types and variations under diverse lighting conditions.
- The above-described problems are addressed and a technical solution achieved in the art by providing a method and system for analyzing at least one food item on a food plate, the method being executed by at least one processor, comprising the steps of receiving a plurality of images of the food plate; receiving a description of the at least one food item on the food plate; extracting a list of food items from the description; classifying and segmenting the at least one food item from the list using color and texture features derived from the plurality of images; and estimating the volume of the classified and segmented at least one food item. The system and method may be further configured for estimating the caloric content of the at least one food item. The description may be at least one of a voice description and a text description. The system and method may be further configured for profiling at least one of the user and meal to include at least one food item not input during the step of receiving a description of the at least one food item on the food plate.
- Classifying and segmenting the at least one food item may further comprise: applying an offline feature-based learning method of different food types to train a plurality of classifiers to recognize individual food items; and applying an online feature-based segmentation and classification method using at least a subset of the food type recognition classifiers trained during offline feature-based learning. Applying an offline feature-based learning method may further comprise: selecting at least three images of the plurality of images, the at least three images capturing the same scene; color normalizing one of the three images; employing an annotation tool is used to identify each food type; and processing the color normalized image to extract color and texture features. Applying an online feature-based segmentation and classification method may further comprise: selecting at least three images of the plurality of images, the at least three images capturing the same scene; color normalizing one of the three images; locating the food plate using a contour based circle detection method; and processing the color normalized image to extract color and texture features. Color normalizing may comprise detecting a color pattern in the scene.
- According to an embodiment of the invention, processing the at least three images to extract color and texture features may further comprise: transforming color features to a CIE L*A*B color space; determining 2D texture features by applying a histogram of orientation gradient (HOG) method; and placing the color features and 2D texture features into bins of histograms in a higher dimensional space. The method may further comprise: representing at least one food type by a cluster of color and texture features in a high-dimensional space using an incremental K-means clustering method; representing at least one food type by texton histograms; and classifying the one food type using an ensemble of boosted SVM classifiers. Applying an online feature-based segmentation and classification method may further comprise: applying a k-nearest neighbors (k-NN) classification method to the extracted color and texture features to each pixel of the color normalized image and assigning at least one label to each pixel; applying a dynamic assembled multi-class classifier to an extracted color and texture feature for each patch of the color normalized image and assigning one label to each patch; and applying an image segmentation technique to obtain a final segmentation of the plate into its constituent food labels.
- According to a preferred embodiment of the invention, the processing the at least three images to extract color and texture features may further comprise: extracting color and texture features using Texton histograms; training a set of one-versus-one classifiers between each pair of foods; and combining color and texture information from the Texton histograms using an Adaboost-based feature selection classifier. Applying an online feature-based segmentation and classification method may further comprise: applying a multi-class classifier to every patch of the three input images to generate a segmentation map; and dynamically assembling a multi-class classifier from a subset of the offline trained pair-wise classifiers to assign a small set of labels to each pixel of the three images.
- Features may be selected for applying a multi-class classifier to every patch of the three input images by employing a bootstrap procedure to sample training data and select features simultaneously. The bootstrap procedure may comprise: randomly sampling a set of training data and computing all features in feature pool; training individual SVM classifiers; applying a 2-fold validation process to evaluate the expected normalized margin for each feature to update the strong classifier; applying a current strong classifier to densely sampled patches in the annotated images, wherein wrongly classified patches are added as new samples, and weights of all training samples are updated; and stopping the training if the number of wrongly classified patches in the training images falls below a predetermined threshold.
- According to an embodiment of the present invention, estimating volume of the classified and segmented at least one food item may further comprise: capturing a set of three 2D images taken at different positions above the food plate with a calibrated image capturing device using an object of known size for 3D scale determination; extracting and matching multiple feature points in each image frame estimating relative camera poses among the three 2D images using the matched feature points; selecting two images from the three 2D images to form a stereo pair and from dense sets of points, determining correspondences between two views of a scene of the two images; performing a 3D reconstruction on the correspondences to generate 3D point clouds of the at least one food item; and estimating the 3D scale and table plane from the reconstructed 3D point cloud to compute the 3D volume of the at least one food item.
- The present invention may be more readily understood from the detailed description of an exemplary embodiment presented below considered in conjunction with the attached drawings and in which like reference numerals refer to similar elements and in which:
-
FIG. 1 is a process flow diagram illustrating exemplary modules/steps for food recognition, according to an embodiment of the present invention; -
FIG. 2 is an exemplary hardware architecture of afood recognition system 30, according to an embodiment of the present invention. -
FIG. 3 shows three images taken by the image capturing device ofFIG. 2 from three different positions; -
FIG. 4A is a process flow diagram illustrating exemplary steps for classifying and segmenting food items using color and texture features employed by the meal content determination module ofFIG. 1 , according to an embodiment of the present invention; -
FIG. 4B is a process flow diagram illustrating exemplary steps for classifying and segmenting food items using color and texture features employed by the meal content determination module ofFIG. 1 , according to a preferred embodiment of the present invention; -
FIG. 5 shows an illustration of the pair-wise classification framework with a set of 10 classes, according to an embodiment of the present invention; -
FIG. 6 is a block diagram of a bootstrap procedure for sampling training data and select features simultaneously for use in the method ofFIG. 4 , according to an embodiment of the present invention; -
FIG. 7 is a process flow diagram illustrating exemplary steps for estimating food volume of a food plate in 3D that has been classified and segmented, according to an embodiment of the present invention; -
FIG. 8A shows a cropped left image of the food plate used in a dense stereo matching step ofFIG. 7 ; -
FIG. 8B shows the corresponding matches between left and right frames, respectively, by a set of horizontal lines using the dense stereo matching step ofFIG. 7 ; -
FIG. 9A displays a top perspective view of a 3D point cloud for an image of the food plate ofFIG. 8A obtained after performing the stereo reconstruction step ofFIG. 7 ; -
FIG. 9B display an example of a 3D point cloud for an image of a food plate obtained after performing stereo reconstruction step ofFIG. 7 ; -
FIG. 10 shows the comparison of sorted pair-wise classification accuracy; -
FIG. 11 is a graph that plots the accuracy of the multi-class classifier obtained during testing of the system ofFIG. 2 ; -
FIG. 12 shows qualitative results of classification and 3D volume estimation obtained during testing of the system ofFIG. 2 ; and -
FIG. 13 shows a plot of error rate per image set for testing the accuracy and repeatability of volume estimation under different capturing conditions obtained during testing of the system ofFIG. 2 . - It is to be understood that the attached drawings are for purposes of illustrating the concepts of the invention and may not be to scale.
-
FIG. 1 is a process flow diagram illustrating exemplary modules/steps for food recognition, according to an embodiment of the present invention.FIG. 2 is an exemplary hardware architecture of afood recognition system 30, according to an embodiment of the present invention. Referring now toFIGS. 1 and 2 , indata capturing module 10, visual and audio and/or text data are captured pertaining to a plate of food. According to a preferred embodiment of the present invention, a plurality of images of a food plate, preferably at least three images, are taken by animage capturing device 32. Theimage capturing device 32 may be, for example, a cell phone or smart phone equipped with a camera, a laptop or desktop computer or workstation equipped with a webcam, or any camera operating in conjunction with a computing platform. In a preferred embodiment, the images are either directly transferred to an image and voice processing server/computer 34 comprising at least one processor directly connected to theimage capturing device 32 via, for example, a USB cable, or remotely to the image and voice processing server/computer 34 over acell network 36 and/or theInternet 38. Indata capturing module 10, according to an embodiment of the present invention, data describing the types of items of food on the food plate may be captured by adescription recognition device 40 for receive a description of items on the food plate from the user in aprocessing step 12. According to an embodiment of the present invention, the description recognition device may be, but is not limited to, a voice recognition device, such as a cell phone or voice phone. Alternatively, thedescription recognition device 40 may be provided with a menu of items that may be present in a meal from which the user chooses, or the user may input food items by inputting text which is recognized by a text recognition device. Theimage capturing device 32 and thedescription recognition device 40 may be integrated in a single device, e.g., a cell phone or smart phone. The image and voice processing server/computer 34 and/or thedescription recognition device 40 may be equipped with automatic speech recognition software. -
FIG. 3 shows threeimages 42 taken by theimage capturing device 32 ofFIG. 2 from three different positions. Theimages 42 may be subject to parallax and substantially different lighting conditions. Thesystem 30 ofFIG. 2 needs to have some guidance with respect to the size of items on a food plate. To this effect, a credit cardsized object 44 may be included in theimages 42 for calibration of 3D size inimage processing module 12. The credit card sized object may be an actual credit card or, for example, an object exhibiting a pattern of black and white squares of known size. The pattern or items located on theobject 44 may be used to establish the relationship between size in image pixels and the actual size offood items 46 on the food plate say, for example, in centimeters. This provides a calibration of pixels per centimeter in theimages 42. - According to an embodiment of the present invention, the automatic speech recognition software in the
voice processing module 14 extracts the list of food from the speech input. Note that the location of the food items on the plate is not specified by the user. Referring again toFIG. 1 , food items identified in thevoice processing module 14 are classified in a mealcontent determination module 16, which makes use of the list of food items provided by the voice/image processing modules - One element of food identification includes plate finding. The list of foods items provided by automatic speech recognition in the
voice processing module 14 is used to initialize food classification in the mealcontent determination module 16. According to an embodiment of the present invention, the food items on the food plate are classified and segmented using color and texture features. Classification and segmentation of food items in the mealcontent determination module 16 is achieved using one or more classifiers known in the art to be described hereinbelow. Inportion estimation module 18, the volume of each of the classified and segmented food items is estimated. - In an optional meal
model creation module 20, the individual segmented food items are reconstructed on a model of the food plate. - In Estimation of
Nutritional Value module 22, the caloric content of the food items of the entire meal may be estimated based on food item types present on the food plate and volume of the food item. In addition to calorie count, other nutritional information may be provided such as, for example, the amount of certain nutrients such as sodium, the amount of carbohydrates versus fat versus protein, etc. - In an optional User
Model Adaption module 24, a user and/or the meal is profiled for potential missing items on the food plate. A user may not identify all of the items on the food plate.Module 24 provides a means of filling in missing items after training thesystem 30 with the food eating habits of a user. For example, a user may always include mashed potatoes in their meal. As a result, thesystem 30 may include probing questions which ask the user at a user interface (not shown) whether the meal also includes items, such as mashed potatoes, that were not originally input in the voice/text recognition module 40 by the user. As another variation, the UserModel Adaption module 24 may statistically assume that certain items not input are, in fact, present in the meal. The UserModel Adaption module 24 may be portion specific, location specific, or even time specific (e.g., a user may be unlikely to dine on a large portion of steak in the morning). - According to an embodiment of the preset invention, plate finding comprises applying the Hough Transform to detect the circular contour of the plate. Finding the plate helps restrict the food classification to the area within the plate. A 3-D depth computation based method may be employed in which the plate is detected using the elevation of the surface of the plate.
- An off-the-shelf speech recognition system may be employed to recognize the list of foods spoken by the end-user into the cell-phone. In one embodiment, speech recognition comprises matching the utterance with a pre-determined list of foods. The
system 30 recognizes words as well as combinations of words. As thesystem 30 is scaled up, speech recognition may be made more flexible by accommodating variations in the food names spoken by the user. If the speech recognition algorithm runs on a remote server, more than sufficient computational resources are available for full-scale speech recognition. Furthermore, since the scope of the speech recognition is limited to names of foods, even with a full-size food name vocabulary, the overall difficulty of the speech recognition task is much less than that of the classic large vocabulary continuous speech recognition problem. -
FIG. 4A is a process flow diagram illustrating exemplary steps for classifying and segmenting food items using color and texture features employed by the mealcontent determination module 16 ofFIG. 1 . Food classification and segmentation is achieved using offline feature-based learning of different food types which ultimately trains a plurality of classifiers to recognize individual food items and online feature-based segmentation and classification using at least a subset of the food type recognition classifiers trained during offline feature-based learning. Inoffline step 50 and again inonline step 60, at least three frames of a plurality of frames are color normalized, the at least three images capturing the same scene. Color differences due to various lighting conditions and picture taking angles occurring in the three frames are synchronized to a single set of colors for each of the frames. To deal with varying lighting conditions, a color pattern is placed in the image for photometric calibration. Fourteen colors (12 from the color pattern and 2 from the checker-board) have been used to solve a 3×3 color transformation matrix using a least squares solution. As texture features may vary with changes in scale, normalization of scale is necessary. For this purpose, a scaling factor is determined to map the checker-board to a predetermined size (75×75 pixel). A color pattern is detected in the scene and one of the three images is color normalized. Atoffline step 52, an annotation tool is used to identify each food type. Annotations may be provided by the user to establish ground truth. Atonline step 62, the plate is located by using a contour based circle detection method proposed in W. Cai, Q. Yu, H. Wang, and J. Zheng “A fast contour-based approach to circle and ellipse detection,” in. 5th IEEE World Congress on Intelligent Control and Automation (WCICA) 2004. The plate is regarded as one label during classification and plate regions are annotated as well in the training set. At bothoffline steps 54 andonline steps 64, the color normalized image is processed to extract color and texture features. Typically the features comprise color features and 2D texture features placed into bins of histograms in a higher dimensional space. The color features are transformed to a CIE L*A*B* color space, wherein the size of the vector of the resulting histogram is: -
Size of feature vector=32 dimensional histogram per channel×3 channels(L,A,B)=96 dimensions - The 2D Texture Features are determined from both extracting HOG features over 3 scales and 4 rotations wherein:
-
Size of feature vector=12 orientation bins×2×2(grid size)=48 dimensions - And from steerable filters over 3 scales and 6 rotations wherein:
-
- Mean and variance of filter response energies is determined over all rotations per scale
- The Size of feature vector=3 scales×2 (mean, variance)×16 bin histogram=96 dimensions
- Histograms of filter outputs are extracted over scale and orientations.
Variations of these features in terms of number of scales and orientations are also incorporated. A key innovation is the use of absolute scale in defining the scale of features by means of a calibration. Since calibration produces an absolute pixels/cm scale, scales are typically chosen in cms for representing the texture of various foods. For instance, scales of 0.5, 1, 2, 4 cms may be used to capture the texture scale of most common foods. Furthermore, an aggregation scale is defined as a multiple of these texture scales. The cms scales are converted to pixels using the calibration. According to an embodiment of the present invention, atoff line step 56, each food type is represented by a cluster of color and texture features in a high-dimensional space using an incremental K-means clustering method. Inoffline step 58, at least one food type is represented by Texton histograms to be described hereinbelow. Food class identification may be performed using an ensemble of boosted SVM classifiers. However, foronline classification step 66, since there may be a large number of food classes to be classified, a k-NN (k-nearest neighbors) classification method is used. The number of clusters chosen for each food type is performed adaptively so that an over-complete set of cluster centers is obtained. During online classification, each pixel's color and texture features are computed and assigned a set of plausible labels using the speech/text input 65 as well as color/texture k-NN classification. A dynamically assembled multi-class classifier may be applied to an extracted color and texture feature for each patch of the color normalized image and one label may be assigned to each patch. Theresult 68 is an assignment of a small set of labels to each pixel.
- Subsequently, an image segmentation technique, such as a Belief Propagation (BP) like technique, may be applied to achieve a final segmentation of the plate into its constituent food labels. For BP, data terms comprising of confidence in the respective color and/or texture feature may be employed. Also, smoothness terms for label continuity may be employed.
-
FIG. 4B is a process flow diagram illustrating exemplary steps for classifying and segmenting food items using color and texture features employed by the mealcontent determination module 16 ofFIG. 1 . According to the preferred embodiment of the present invention ofFIG. 4B , offline and online feature extraction steps 54 and 64, respectively, offline K-means clustering step 56,offline classification step 58, and on-line classification step 66 ofFIG. 4A may be replaced by offlinefeature extraction step 54′ (a corresponding online feature extraction step is not needed),offline classification step 58′ andonline classification step 66′ ofFIG. 4B . The task of food recognition is formulated in steps as a multi-class classification problem. In offlinefeature extraction step 54′, features are extracted using Texton histograms. Inoffline classification step 58′, multi-class recognition problem may be simplified by making use of the user providing a candidate food type set 65 acquired during speech recognition as described above. In order to make full use of this additional cue, a set of one-versus-one classifiers are trained between each pair of foods. A segmentation map is generated by applying a multi-class classifier densely (i.e., every patch) to an input image. An Adaboost-based feature selection classifier is adapted to combine color and texture information to achieve an acceptable food type recognition rate over a large number of food types. Inonline classification step 66′ based on these offline trained pair-wise classifiers, a dynamically assembled trained classifier is created according to the candidate set on the fly to assign a small set of labels to each pixel. - Suppose there exist N classes of food {fi: i=1, . . . , N}, then all the pair-wise classifiers may be represented as C={Cij: i, j∈[1, N], i<j}. The total number of classifiers, |C|, is N×(N−1)/2. For a set of K candidates, K×(K−1)/2 pair-wise classifiers are selected to assemble a K-class classifier. The dominant label assigned by the selected pair-wise classifiers is the output of the K-class classification. If there is no unanimity among the K pair-wise classifiers corresponding to a food type, then the final output is set to unknown.
FIG. 5 shows an illustration of the pair-wise classification framework with a set of 10 classes. The upper triangular matrix contains 45 offline trained classifiers. For example, if 5 classes are chosen as candidates by the user, then 10 pair-wise classifiers may be assembled to form a 5-class classifier. If 4 out of 10 classifiers report the same label, this label is reported as the final label, otherwise an unknown label is reported by the 5-class classifier. - The advantages of this framework are two-fold. First, computation cost is reduced during the testing phase. Second, compared with one-versus-all classifiers, this framework avoids N imbalance in training samples (a few positive samples versus a large number of negative samples). Another strength of this framework is its extendibility. Since there are a large number of food types, users of the
system 30 ofFIG. 2 may incrementally update existing classes with new instances and add new food types without re-training classifiers from scratch. This pair-wise framework is easy to adapt to new classes and to new instances. If there exists N pre-trained classes, then updating a class may be accomplished by re-training (N−1) classifiers in the upper triangular matrix; adding a new class, named fN+1, is equivalent to adding a new column (N) of classifiers {Ci,N+1: i=1, . . . , N}. - To compute a label map (i.e., labels for items on a food plate), classifiers are applied densely (every patch) on the color and scale normalized images. To train such a classifier, the training set is manually annotated to obtain segmentation, in the form of label masks, of the food. Texton histograms are used as features for classification, which essentially translate to a bag-of-words. There are many approaches that have been proposed to create textons, such as spatial-frequency based textons as described in M. Varma and A. Zisserman, “Classify images of materials: Achieving viewpoint and illumination independence,” in ECCV, pages 255-271, 2002 (hereinafter “Varma1”), MRF textons as described in M. Varma and A. Zisserman, “Texture classification: Are filter banks necessary?” In CVPR, pages 691-698, 2003 (hereinafter “Varma2”), and gradient orientation based textons as described in D. Lowe, “Distinctive image features from scale-invariant keypoints,” IJCV, pages 91-110, 2004. A detailed survey and comparison of local image descriptors may be found in K. Mikolajczyk and C. Schmid, “A performance evaluation of local descriptors,” PAMI, pages 1615-1630, 2005.
- It is important to choose the right texton as it directly determines the discriminative power of texton histograms. The current features used in the
system 30 include color (RGB and LAB) neighborhood features as described in Varma1 and Maximum Response (MR) features as described in Varma2. The color neighborhood feature is a vector that concatenates color pixels within an L×L patch. Note that for the case L=1 this feature is close to a color histogram. An MR feature is computed using a set of edge, bar, and block filters along 6 orientations and 3 scales. Each feature comprises eight dimensions by taking a maximum along each orientation as described in Varma2. Note that when the convolution window is large, convolution is directly applied to the image instead of patches. Filter responses are computed and then a feature vector is formed according to a sampled patch. Both color neighborhood and MR features may be computed densely in an image since the computational cost is relatively low. Moreover, these two types of features contain complementary information: the former contains color information but cannot carry edge information at a large scale, which is represented in the latter MR features; the latter MR features do not encode color information, which is useful to separate foods. It has been observed that by using only one type of feature at one scale a satisfactory result cannot be achieved over all pair-wise classifiers. As a result, feature selection may be used to create a strong classifier from a set of weak classifiers. - A pair of foods may be more separable using some features at a particular scale than using other features at other scales. In training a pair-wise classifier, all possible types and scales of features may be choose and concatenated into one feature vector. This, however, puts too much burden on the classifier by confusing it with non-discriminative features. Moreover, this is not computationally efficient. Instead, a rich set of local feature options (color, texture, scale) may be created and a process of feature selection may be employed to automatically determine the best combination of heterogeneous features. The types and scales of features used in current system are shown in Table 1.
-
TABLE 1 Features options Type Scale Color(RGB/LAB) Neighborhood (See Varma1) 1, 3, 5, 7 Maximum Responses (See Varma2) 0.5, 1, 2 - The feature selection algorithm is based on Adaboost as described in R. E. Schapire, Y. Freund, P. Bartlett, and W. S. Lee, “Boosting the margin: A new explanation for the effectiveness of voting methods,” The Annals of Statistics, pages 1651-1686, 1998, which is an iterative approach for building strong classifiers out of a collection of “weak” classifiers. Each weak classifier corresponds to one type of texton histogram. An χ2 kernel SVM is adopted to train the weak classifier using one feature in the feature pool. A comparison of different kernels in J. Zhang, M. Marszalek, S. Lazebnik, and C. Schmid, “Local features and kernels for classification of texture and object categories: A comprehensive study,” IJCV, pages 213-238, 2007, shows that χ2 kernels outperform the rest.
- A feature set {f1, . . . , fn} is denoted by F. In such circumstances, a strong classifier based on a subset of features by F⊂F may be obtained by linear combination of selected weak SVM classifiers, h: X→R,
-
- where
-
- and εf is the weighed error rate of the weak classifier f1. For a sample x, denote its true class label by y(=±1). The classification margin of h on x is defined by y×h(x). The classification margin represents the discriminative power of the classifier. Larger margins imply better generalization power. Adaboost is an approach to iteratively select the feature in the feature pool which has the largest margin according to current distribution (weights) of samples.
-
- where Hk is the strong classifier learned in the kth round and M(•) is the expected margin on X.
- As each h is a SVM, this margin may be evaluated by N-fold validation (in our case, we use N=2). Instead of comparing the absolute margin of each SVM, a normalized margin is adopted, as
-
- where PhP denotes the number of support vectors. This criterion actually measures the discriminative power per support vector. This criterion avoids choosing a large-margin weak classifier that is built with many support vectors and possibly overfits the training data. Also, this criterion tends to produce a smaller number of support vectors to ensure low complexity.
- Another issue in the present invention is how to make full use of training data. Given annotated training images and a patch scale, a large number of patches may be extracted by rotating and shifting the sampling windows. Instead of using a fixed number of training samples or using all possible training patches, a bootstrap procedure is employed as shown in
FIG. 6 to sample training data and select features simultaneously. Initially, atstep 70, a set of training data is randomly sampled and all features in feature pool are computed. Atstep 72, individual SVM classifiers are trained. Atstep 74, a 2-fold validation process is employed to evaluate the expected normalized margin for each feature and the best one is chosen to update the strong classifier with weighted classification error instep 76. The current strong classifier is applied to densely sampled patches in the annotated images, wrongly classified patches (plus the ones close to the decision boundary) are added as new samples, and weights of all training samples are updated. Note that instep 69 training images in the LAB color space are perturbed before bootstrapping. The training is stopped if the number of wrongly classified patches in the training images falls below a predetermined threshold. -
FIG. 7 is a process flow diagram illustrating exemplary steps for estimating food volume of a food plate in 3D that has been classified and segmented. In order to estimate the volume of food items on a user's plate, atstep 80, a set of three 2D images is taken at different positions above the plate with a calibrated image capturing device (e.g., a cell phone camera) using an object of known size for 3D scale determination. Atstep 82, multiple feature points in each image frame are extracted and matches between the three 2D images. Atstep 84, using the matched feature points, the relative camera poses among the three 2D images are estimated. In a densestereo matching step 86, two images from the three 2D images are selected to form a stereo pair and from dense sets of points, correspondences between the two views of a scene of the two images are found. Instep 88, a 3D reconstruction is carried out on the correspondences to generate 3D point clouds of the food. Finally, from the reconstructed 3D point cloud, both the 3D scale (step 92) and table plane are estimated to compute the 3D volume of each food item (step 94). - According to an embodiment of the present invention, and referring again to step 82, the multiple feature points in each of the three 2D images are extracted matches between images using Harris corners, as described in C. Harris and M. Stephens, “A combined corner and edge detector,” in the 4th Alvey Vision Conference, 1988. However, any other feature which describes an image point in a distinctive manner may be used. Each feature correspondence establishes a feature track, which lasts as long as it is matched across the images. These feature tracks are later sent into the
pose estimation step 84 which is carried out using a preemptive RANSAC-based method as described in D. Nister, O. Naroditsky, and J. Bergen, “Visual odometry,” in CVPR, 2004, as explained in more detail hereinbelow. - The preemptive RANSAC algorithm randomly selects different sets of 5-point correspondences over three frames such that N number of pose hypotheses (by default N=500) are generated using a 5-point algorithm. Here, each pose hypothesis comprises the pose of the second and third view with respect to the first view. Then, starting with all of the hypotheses, each one is evaluated on chunks of M data points based on trifocal Sampson error (by default M=100), every time dropping out half of the least scoring hypotheses. Thus, initially, 500 pose hypotheses are proposed, all of which are evaluated on a subset of 100-point correspondences. Then the 500 pose hypotheses are sorted according to their scores on the subset of 100-point correspondences and the bottom half is removed. In the next step, another set of 100 data points is selected on which the remaining 250 hypotheses are evaluated and the least scoring half are pruned. This process continues until a single best-scoring pose hypothesis remains.
- In the next step, the best pose at the end of the preemptive RANSAC routine is passed to a pose refinement step where iterative minimization of a robust cost function (derived from Cauchy distribution) of the re-projection errors is performed through Levenberg-Marquardt method as described in R. Hartley and A. Zisserman, “Multiple View Geometry in Computer Vision,” Cambridge University Press, 2000, pp. 120-122.
- Using the above proposed algorithm, camera poses are estimated over three views such that poses for the second and third view are with respect to the camera coordinate frame in the first view. In order to stitch these poses, the poses are placed in the coordinate system of the first camera position corresponding to the first frame in the image sequence. At this point, the scale factor for the new pose-set (poses corresponding to the second and third views in the current triple) is also estimated with another RANSAC scheme.
- Once the relative camera poses between the image frames have been estimated, in a dense
stereo matching step 86, two images from the three 2D images are selected to form a stereo pair and from dense sets of points, correspondences between the two views of a scene of the two images are determined. For each pixel in the left image, its corresponding pixel in the right image is searched using a hierarchal pyramid matching scheme. Once the left-right correspondence is found, instep 88, using the intrinsic parameters of the pre-calibrated camera, the left-right correspondence match is projected in 3D using triangulation. At this stage, any bad matches are filtered out by validating them against the epipolar constraint. To gain speed, the reconstruction process is carried out for all non-zero pixels in the segmentation map provided by the food classification stage.FIGS. 8A and 8B show the corresponding matches between left and right frames, respectively, by a set ofhorizontal lines 100. - Referring again to
FIG. 7 , after thepose estimation step 84, there is still a scale ambiguity in the final pose of the three 2D frames. In order to recover a global scale factor, an object with known dimensions is placed and captured along with the plate of food in the image. For simplicity, according to an embodiment of the present invention, a checker-board may be employed. In order to compute 3D scale, each corner of the checker-board in an image is detected followed by its reconstruction to obtain corresponding 3D coordinates. The size of each checker-board square is determined in 3D from its respective corners. Let dRef be the real size of checker-board as measured by ground truth and dEst be its size as obtained by estimation in 3D. Then, the 3D scale (S) is computed usingequation 3. In the present embodiment, a 3×3 checker-board may be used, with dRef=3.14 cms. -
S=d Ref /d Est (3) - Once the 3D scale is computed using the checker-board, an overall scale correction is made to all the camera poses over the set of frames and the frames are mapped to a common coordinate system. Following stereo reconstruction, a dense 3D point cloud for all points on the plate is obtained.
FIGS. 8A and 8B display an example of a 3D point cloud for an image of a food plate. Since the volume of each food item needs to be measured with respect to a reference surface, estimation of the table plane is carried out as a pre-requisite step. By inspection of the image, a person skilled in the art would appreciate that, apart from pixels corresponding to food on the plate, most pixels lie on the table plane. Hence, table estimation is performed by employing RANSAC to fit a 3D plane equation on feature points earlier used for camera pose estimation. To obtain better accuracy, points falling on the plate are removed for the purpose of plane fitting by using the boundaries obtained from the plate detection step. Once the table plane has been estimated, it is used to slice the entire point cloud into two portions such that only 3D points above the plane are considered for the purpose of volume estimation. - Referring again to
FIG. 7 , thevolume estimation step 92 is carried out in two sub-steps. First, Delaunay triangulation is performed to fit the surface of food. Second, total volume of the food (Vtotal) is calculated as a sum of individual volumes (Vi) for each Delaunay triangles obtained from the previous step.Equation 4 shows computation of total food volume where K is the total number of triangles. -
- One of the main tasks of the present invention is to report volumes of each individual food item on a user's plate. This is done by using the binary label map obtained after food recognition. The label map for each food item consists of non-zero pixels that have been identified as belonging to the food item of interest and zero otherwise. Using this map, a subset of the 3D point cloud is selected that corresponds to reconstruction of a particular food label that is then feed it into the volume estimation process. This step is repeated for all food items on the plate to compute their respective volumes.
- Experiments were carried out to test the accuracy of certain embodiments of the present invention. In order to standardize analysis of various foods, the USDA Food and Nutrient Database for Dietary Studies (FNDDS) was consulted, which contains more than 7,000 foods along with the information such as, typical portion size and nutrient value. 400 sets of images containing 150 commonly occurring food types in the FNDDS were collected. This data was used to train classifiers. An independently collected data set with 26 types of foods was used to evaluate the recognition accuracy. N (in this case, N=500) patches were randomly sampled from images of each type of food and the accuracy of classifiers trained in different ways was evaluated as follows:
-
- Using a single MR feature (σx
1 =0.5); - Using single RGB neighborhood features (at 3×3 scale);
- Using combined features with fixed number of training samples per food label;
- Using feature selection in the proposed bootstrap framework.
- Using a single MR feature (σx
- For comparison, all pair-wise classifiers were trained (13×25=325) and classification accuracy was sorted. As each pair-wise classifier ci,j was evaluated over 2N patches (N patches in label i and N patches in label j), the pair-wise classification accuracy is the ratio of correct instances over 2N.
FIG. 10 shows the comparison of sorted pair-wise classification accuracy. By applying the feature selection in the bootstrap procedure, a significant improvement was achieved over using a single feature and using a fixed number of training samples. - In order to evaluate the multi-class classifiers assembled online based on user input, K confusing labels were randomly added to each ground truth label in the test set. Hence, the multi-class classifier had K+1 candidates. The accuracy of the multi-class classifier is shown in
FIG. 11 . As can be seen inFIG. 11 , accuracy drops as the number of candidates increases. The larger the number of candidates, the more likely the confusion between them. However, the number of foods in a meal is rarely greater than 6, for which about a 90% accuracy was achieved. - Qualitative results of classification and 3D volume estimation are shown in
FIG. 12 (Table 2): the first column shows the images after scale and color normalization; the second column shows the classification results and the last column shows the reconstructed 3D surface obtained using Delaunay triangulation and the estimated table plane, which are used for computing the volume. Table 3 shows the quantitative evaluation of these sets. In the system of the present invention, volume is returned in milliliter units. This value may be converted to calories by indexing into the FNDDS. -
TABLE 3 Quantitative classification and 3D volume results Ground truth Estimate Error Set # Food (in ml) (in ml) (%) 1 Broccoli 150 143.5 4.3 Carrots 120 112.3 6.4 2 Orange 195 189.4 2.9 Bagel 300 310.5 3.5 3 Fries 200 194.8 2.6 Steak 190 203.9 7.3 Broccoli 180 186.3 3.5 4 Spinach 160 151.2 5.5 Cucumber 100 98.2 1.5 Olives 100 104.8 4.8 Broccoli 120 114.2 4.8 Peppers 80 82.7 3.4 5 Olives 100 98.4 1.6 Carrots 90 82.7 8.1 Peas 120 123.8 3.2 Chickpeas 100 103.1 3.1 Cucumber 140 144.2 3.0 Peppers 90 84.1 6.6 6 Chicken 130 121.2 6.8 Fries 150 133.6 10.9 - To test the accuracy and repeatability of volume estimation under different capturing conditions, an object with a known ground truth volume is given as input to the system. For this evaluation, 35 image sets of the object were captured taken at different viewpoints and heights.
FIG. 13 shows a plot of error rate per image set. The average error in volume is 5.75 (±3.75) % over all the sets. - The experimental system was run on a Intel Xeon workstation with 3 GHz CPU and 4 GB of RAM. The total turn-around time was 52 seconds (19 seconds for classification and 33 seconds for dense stereo reconstruction and volume estimation on a 1600×1200 pixel image). The experimental system was not optimized and ran on a single core.
- It is to be understood that the exemplary embodiments are merely illustrative of the invention and that many variations of the above-described embodiments may be devised by one skilled in the art without departing from the scope of the invention. It is therefore intended that all such variations be included within the scope of the following claims and their equivalents.
Claims (20)
1. A method for analyzing one or more food items on a food plate, the method being executed by at least one processor, comprising the steps of:
extracting a list of food items from a description of the one or more food items on the foot plate; and
classifying and segmenting the one or more food item from the list to identify the one or more food items using color and texture features derived from one or more images of the food plate by: and
applying an online feature-based segmentation and classification method using one or more food type recognition classifiers trained by offline feature-based learning.
2. The method of claim 1 , further comprising the step of estimating the caloric content of the one or more food item.
3. The method of claim 1 , wherein the description is one or more of a voice description and a text description.
4. The method of claim 1 , wherein applying an offline feature-based learning method further comprises the steps of:
selecting three or more images which capture the same scene of the food plate by receiving a plurality of images of the food plate in addition to the one or more images;
color normalizing one or more of the three or more images;
employing an annotation tool used to identify each food type; and
processing the color normalized image to extract color and texture features of each of the food items.
5. The method of claim 4 , wherein the step of color normalizing comprises detecting a color pattern in the scene.
6. The method of claim 4 , wherein the step of processing the color normalized image to extract color and texture features of each of the food items further comprises the steps of:
transforming color features to a CIE L*A*B color space;
determining 2D texture features by applying a histogram of orientation gradient (HOG) method; and
placing the color features and 2D texture features into bins of histograms in a higher dimensional space.
7. The method of claim 4 , wherein the step of processing the color normalized image to extract color and texture features further comprises the steps of:
extracting color and texture features using Texton histograms;
training a set of one-versus-one classifiers between each pair of foods; and
combining color and texture information from the Texton histograms using an Adaboost-based feature selection classifier.
8. The method of claim 1 , wherein the step of applying an online feature-based segmentation and classification method further comprises the steps of:
selecting three or more images of a plurality of images received in addition to the one or more images of the food plate, the three or more images capturing the same scene of the food plate;
color normalizing the three or more images;
locating the food plate using a contour based circle detection method; and
processing the color normalized image to extract color and texture features.
9. The method of claim 8 , wherein the step of color normalizing comprises detecting a color pattern in the scene.
10. The method of claim 8 , further comprising the steps of:
representing one or more food types by a cluster of color and texture features in a high-dimensional space using an incremental K-means clustering method;
representing another one or more food types by Texton histograms; and
classifying the one or more food type using an ensemble of boosted SVM classifiers.
11. The method of claim 8 , wherein the step of applying an online feature-based segmentation and classification method further comprises the steps of:
applying a k-nearest neighbors (k-NN) classification method to the extracted color and texture features to each pixel of the color normalized image and assigning at least one label to each pixel;
applying a dynamic assembled multi-class classifier to an extracted color and texture feature for each patch of the color normalized image and assigning one label to each patch; and
applying an image segmentation technique to obtain a final segmentation of the plate into its constituent food labels.
12. The method of claim 8 , wherein the step of applying an online feature-based segmentation and classification method further comprises the steps of:
applying a multi-class classifier to every patch of the three input images to generate a segmentation map; and
dynamically assembling a multi-class classifier from a subset of the offline trained pair-wise classifiers to assign a small set of labels to each pixel of the three images.
13. The method of claim 12 , wherein features are selected for applying a multi-class classifier to every patch of the three input images by employing a bootstrap procedure to sample training data and select features simultaneously.
14. The method of claim 13 , wherein the bootstrap procedure comprises the steps of: randomly sampling a set of training data and computing all features in feature pool;
training individual SVM classifiers;
applying a 2-fold validation process to evaluate the expected normalized margin for each feature to update the strong classifier;
applying a current strong classifier to densely sampled patches in the annotated images, wherein wrongly classified patches are added as new samples, and weights of all training samples are updated; and
stopping the training if the number of wrongly classified patches in the training images falls below a predetermined threshold.
15. The method of claim 1 , wherein the step of estimating volume of the classified and segmented one or more food items further comprises the steps of:
capturing a set three or more 2D images from a plurality of images received in addition to the one or more images taken at different positions above the food plate with a calibrated image capturing device using an object of known size for 3D scale determination;
extracting and matching multiple feature points in each image frame;
estimating camera poses among the three or more images using the matched feature points;
selecting two or more images from the three or more images;
determining correspondences between the two or more images selected from at least the three or more images;
performing a 3D reconstruction on the correspondences and determining a 3D scale based on the object of known size to generate 3D point cloud;
estimating one or more surfaces of the one or more food items above the food plate based on at least the reconstructed 3D point cloud: and
estimating the volume of the one or more food items based on the one or more surfaces.
16. A system for analyzing one or more food items on a food plate, comprising:
a processor for:
classifying and segmenting the one or more food items using one or more features of the one or more food items derived from one or more images, by applying an online feature-based segmentation and classification method using one or more food type recognition classifiers trained during offline feature-based learning; and
estimating the volume of the classified and segmented one or more food items based on determining correspondences between images of the one or more images containing the one or more food items.
17. The system of claim 16 , wherein the processor further:
receives a description from a device describing the one or more food items on the food plate,
wherein the description device is at least one of a voice recognition device and a text recognition device and also supplies the one or more images.
18. The system of claim 17 , wherein the image capturing device is one of a cell phone or smart phone equipped with a camera, a laptop or desktop computer or workstation equipped with a webcam, or a camera operating in conjunction with a computing platform.
19. The system of claim 17 , wherein the processor is integrated into a voice processing computer, which is one of: directly connected to the image capturing device and remotely over a cell network and/or the Internet.
20. The system of claim 17 , wherein the processor estimates the caloric content of the one or more food items.
Priority Applications (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
US13/849,222 US20130260345A1 (en) | 2009-01-07 | 2013-03-22 | Food recognition using visual analysis and speech recognition |
US15/818,528 US20180075774A1 (en) | 2009-01-07 | 2017-11-20 | Food recognition using visual analysis and speech recognition |
Applications Claiming Priority (3)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
US14308109P | 2009-01-07 | 2009-01-07 | |
US12/683,124 US8439683B2 (en) | 2009-01-07 | 2010-01-06 | Food recognition using visual analysis and speech recognition |
US13/849,222 US20130260345A1 (en) | 2009-01-07 | 2013-03-22 | Food recognition using visual analysis and speech recognition |
Related Parent Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
US12/683,124 Continuation US8439683B2 (en) | 2009-01-07 | 2010-01-06 | Food recognition using visual analysis and speech recognition |
Related Child Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
US15/818,528 Continuation US20180075774A1 (en) | 2009-01-07 | 2017-11-20 | Food recognition using visual analysis and speech recognition |
Publications (1)
Publication Number | Publication Date |
---|---|
US20130260345A1 true US20130260345A1 (en) | 2013-10-03 |
Family
ID=42311930
Family Applications (3)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
US12/683,124 Active 2031-07-04 US8439683B2 (en) | 2009-01-07 | 2010-01-06 | Food recognition using visual analysis and speech recognition |
US13/849,222 Abandoned US20130260345A1 (en) | 2009-01-07 | 2013-03-22 | Food recognition using visual analysis and speech recognition |
US15/818,528 Abandoned US20180075774A1 (en) | 2009-01-07 | 2017-11-20 | Food recognition using visual analysis and speech recognition |
Family Applications Before (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
US12/683,124 Active 2031-07-04 US8439683B2 (en) | 2009-01-07 | 2010-01-06 | Food recognition using visual analysis and speech recognition |
Family Applications After (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
US15/818,528 Abandoned US20180075774A1 (en) | 2009-01-07 | 2017-11-20 | Food recognition using visual analysis and speech recognition |
Country Status (1)
Country | Link |
---|---|
US (3) | US8439683B2 (en) |
Cited By (15)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20130170714A1 (en) * | 2010-05-31 | 2013-07-04 | The University Of Tokyo | Information processing device |
US9042596B2 (en) | 2012-06-14 | 2015-05-26 | Medibotics Llc | Willpower watch (TM)—a wearable food consumption monitor |
US9254099B2 (en) | 2013-05-23 | 2016-02-09 | Medibotics Llc | Smart watch and food-imaging member for monitoring food consumption |
US9442100B2 (en) | 2013-12-18 | 2016-09-13 | Medibotics Llc | Caloric intake measuring system using spectroscopic and 3D imaging analysis |
US9529385B2 (en) | 2013-05-23 | 2016-12-27 | Medibotics Llc | Smart watch and human-to-computer interface for monitoring food consumption |
US9536449B2 (en) | 2013-05-23 | 2017-01-03 | Medibotics Llc | Smart watch and food utensil for monitoring food consumption |
US9734426B2 (en) | 2014-09-03 | 2017-08-15 | Sri International | Automated food recognition and nutritional estimation with a personal mobile electronic device |
CN107194418A (en) * | 2017-05-10 | 2017-09-22 | 中国科学院合肥物质科学研究院 | A kind of Aphids in Rice Field detection method based on confrontation feature learning |
US9972080B2 (en) | 2016-01-14 | 2018-05-15 | Electronics And Telecommunications Research Institute | Apparatus and method for food search service |
US9977980B2 (en) * | 2014-02-12 | 2018-05-22 | Microsoft Technology Licensing, Llc | Food logging from images |
US10521903B2 (en) | 2015-11-25 | 2019-12-31 | Samsung Electronics Co., Ltd. | User terminal apparatus and control method thereof |
US10772559B2 (en) | 2012-06-14 | 2020-09-15 | Medibotics Llc | Wearable food consumption monitor |
US10943398B2 (en) | 2016-07-15 | 2021-03-09 | Samsung Electronics Co., Ltd. | Augmented reality device and operation thereof |
US11069045B2 (en) | 2018-08-03 | 2021-07-20 | International Business Machines Corporation | Food preparation assistance using image processing |
US11272810B2 (en) | 2014-03-14 | 2022-03-15 | Spectrum Brands, Inc. | Food preparation appliance for use with a remote communication device |
Families Citing this family (104)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US10698939B2 (en) | 2005-10-26 | 2020-06-30 | Cortica Ltd | System and method for customizing images |
US10387914B2 (en) | 2005-10-26 | 2019-08-20 | Cortica, Ltd. | Method for identification of multimedia content elements and adding advertising content respective thereof |
US11386139B2 (en) | 2005-10-26 | 2022-07-12 | Cortica Ltd. | System and method for generating analytics for entities depicted in multimedia content |
US10742340B2 (en) | 2005-10-26 | 2020-08-11 | Cortica Ltd. | System and method for identifying the context of multimedia content elements displayed in a web-page and providing contextual filters respective thereto |
US9767143B2 (en) | 2005-10-26 | 2017-09-19 | Cortica, Ltd. | System and method for caching of concept structures |
US9191626B2 (en) | 2005-10-26 | 2015-11-17 | Cortica, Ltd. | System and methods thereof for visual analysis of an image on a web-page and matching an advertisement thereto |
US9558449B2 (en) | 2005-10-26 | 2017-01-31 | Cortica, Ltd. | System and method for identifying a target area in a multimedia content element |
US10193990B2 (en) | 2005-10-26 | 2019-01-29 | Cortica Ltd. | System and method for creating user profiles based on multimedia content |
US9466068B2 (en) | 2005-10-26 | 2016-10-11 | Cortica, Ltd. | System and method for determining a pupillary response to a multimedia data element |
US11620327B2 (en) | 2005-10-26 | 2023-04-04 | Cortica Ltd | System and method for determining a contextual insight and generating an interface with recommendations based thereon |
US8312031B2 (en) | 2005-10-26 | 2012-11-13 | Cortica Ltd. | System and method for generation of complex signatures for multimedia data content |
US11032017B2 (en) | 2005-10-26 | 2021-06-08 | Cortica, Ltd. | System and method for identifying the context of multimedia content elements |
US10380623B2 (en) | 2005-10-26 | 2019-08-13 | Cortica, Ltd. | System and method for generating an advertisement effectiveness performance score |
US9953032B2 (en) | 2005-10-26 | 2018-04-24 | Cortica, Ltd. | System and method for characterization of multimedia content signals using cores of a natural liquid architecture system |
US10614626B2 (en) | 2005-10-26 | 2020-04-07 | Cortica Ltd. | System and method for providing augmented reality challenges |
US10380267B2 (en) | 2005-10-26 | 2019-08-13 | Cortica, Ltd. | System and method for tagging multimedia content elements |
US9384196B2 (en) | 2005-10-26 | 2016-07-05 | Cortica, Ltd. | Signature generation for multimedia deep-content-classification by a large-scale matching system and method thereof |
US10380164B2 (en) | 2005-10-26 | 2019-08-13 | Cortica, Ltd. | System and method for using on-image gestures and multimedia content elements as search queries |
US10635640B2 (en) | 2005-10-26 | 2020-04-28 | Cortica, Ltd. | System and method for enriching a concept database |
US10372746B2 (en) | 2005-10-26 | 2019-08-06 | Cortica, Ltd. | System and method for searching applications using multimedia content elements |
US9218606B2 (en) | 2005-10-26 | 2015-12-22 | Cortica, Ltd. | System and method for brand monitoring and trend analysis based on deep-content-classification |
US10621988B2 (en) | 2005-10-26 | 2020-04-14 | Cortica Ltd | System and method for speech to text translation using cores of a natural liquid architecture system |
US11403336B2 (en) | 2005-10-26 | 2022-08-02 | Cortica Ltd. | System and method for removing contextually identical multimedia content elements |
US10607355B2 (en) | 2005-10-26 | 2020-03-31 | Cortica, Ltd. | Method and system for determining the dimensions of an object shown in a multimedia content item |
US9477658B2 (en) | 2005-10-26 | 2016-10-25 | Cortica, Ltd. | Systems and method for speech to speech translation using cores of a natural liquid architecture system |
US9639532B2 (en) | 2005-10-26 | 2017-05-02 | Cortica, Ltd. | Context-based analysis of multimedia content items using signatures of multimedia elements and matching concepts |
US10691642B2 (en) | 2005-10-26 | 2020-06-23 | Cortica Ltd | System and method for enriching a concept database with homogenous concepts |
US10776585B2 (en) | 2005-10-26 | 2020-09-15 | Cortica, Ltd. | System and method for recognizing characters in multimedia content |
US11216498B2 (en) | 2005-10-26 | 2022-01-04 | Cortica, Ltd. | System and method for generating signatures to three-dimensional multimedia data elements |
US20140093844A1 (en) * | 2005-10-26 | 2014-04-03 | Cortica, Ltd. | Method for identification of food ingredients in multimedia content |
US8266185B2 (en) | 2005-10-26 | 2012-09-11 | Cortica Ltd. | System and methods thereof for generation of searchable structures respective of multimedia data content |
US9646005B2 (en) | 2005-10-26 | 2017-05-09 | Cortica, Ltd. | System and method for creating a database of multimedia content elements assigned to users |
US11003706B2 (en) | 2005-10-26 | 2021-05-11 | Cortica Ltd | System and methods for determining access permissions on personalized clusters of multimedia content elements |
US11019161B2 (en) | 2005-10-26 | 2021-05-25 | Cortica, Ltd. | System and method for profiling users interest based on multimedia content analysis |
US9529984B2 (en) | 2005-10-26 | 2016-12-27 | Cortica, Ltd. | System and method for verification of user identification based on multimedia content elements |
US11604847B2 (en) | 2005-10-26 | 2023-03-14 | Cortica Ltd. | System and method for overlaying content on a multimedia content element based on user interest |
US10180942B2 (en) | 2005-10-26 | 2019-01-15 | Cortica Ltd. | System and method for generation of concept structures based on sub-concepts |
US10585934B2 (en) | 2005-10-26 | 2020-03-10 | Cortica Ltd. | Method and system for populating a concept database with respect to user identifiers |
US10191976B2 (en) | 2005-10-26 | 2019-01-29 | Cortica, Ltd. | System and method of detecting common patterns within unstructured data elements retrieved from big data sources |
US11361014B2 (en) | 2005-10-26 | 2022-06-14 | Cortica Ltd. | System and method for completing a user profile |
US8326775B2 (en) | 2005-10-26 | 2012-12-04 | Cortica Ltd. | Signature generation for multimedia deep-content-classification by a large-scale matching system and method thereof |
US10535192B2 (en) | 2005-10-26 | 2020-01-14 | Cortica Ltd. | System and method for generating a customized augmented reality environment to a user |
US10360253B2 (en) | 2005-10-26 | 2019-07-23 | Cortica, Ltd. | Systems and methods for generation of searchable structures respective of multimedia data content |
US8818916B2 (en) | 2005-10-26 | 2014-08-26 | Cortica, Ltd. | System and method for linking multimedia data elements to web pages |
US9031999B2 (en) | 2005-10-26 | 2015-05-12 | Cortica, Ltd. | System and methods for generation of a concept based database |
US10848590B2 (en) | 2005-10-26 | 2020-11-24 | Cortica Ltd | System and method for determining a contextual insight and providing recommendations based thereon |
US9372940B2 (en) | 2005-10-26 | 2016-06-21 | Cortica, Ltd. | Apparatus and method for determining user attention using a deep-content-classification (DCC) system |
US9489431B2 (en) | 2005-10-26 | 2016-11-08 | Cortica, Ltd. | System and method for distributed search-by-content |
US10949773B2 (en) | 2005-10-26 | 2021-03-16 | Cortica, Ltd. | System and methods thereof for recommending tags for multimedia content elements based on context |
US10733326B2 (en) | 2006-10-26 | 2020-08-04 | Cortica Ltd. | System and method for identification of inappropriate multimedia content |
US8363913B2 (en) * | 2008-09-05 | 2013-01-29 | Purdue Research Foundation | Dietary assessment system and method |
EP2629211A1 (en) | 2009-08-21 | 2013-08-21 | Mikko Kalervo Väänänen | Method and means for data searching and language translation |
US9197736B2 (en) | 2009-12-31 | 2015-11-24 | Digimarc Corporation | Intuitive computing methods and systems |
US10224036B2 (en) * | 2010-10-05 | 2019-03-05 | Infraware, Inc. | Automated identification of verbal records using boosted classifiers to improve a textual transcript |
US9412050B2 (en) * | 2010-10-12 | 2016-08-09 | Ncr Corporation | Produce recognition method |
US8625889B2 (en) | 2010-12-30 | 2014-01-07 | Samsung Electronics Co., Ltd. | System for food recognition method using portable devices having digital cameras |
US20120183932A1 (en) * | 2011-01-14 | 2012-07-19 | International Business Machines Corporation | Location-Aware Nutrition Management |
WO2012115297A1 (en) * | 2011-02-25 | 2012-08-30 | Lg Electronics Inc. | Analysis of food items captured in digital images |
JP2013025728A (en) * | 2011-07-26 | 2013-02-04 | Sony Corp | Information processing apparatus, information processing method, and program |
JP2013054541A (en) * | 2011-09-05 | 2013-03-21 | Sony Corp | Information processor, information processing method, and program |
US8761922B2 (en) * | 2011-10-10 | 2014-06-24 | Jonathan P Dekar | Method and apparatus for monitoring food consumption by an individual |
US10006896B2 (en) | 2011-11-14 | 2018-06-26 | University of Pittsburgh—of the Commonwealth System of Higher Education | Method, apparatus and system for food intake and physical activity assessment |
JP6069826B2 (en) * | 2011-11-25 | 2017-02-01 | ソニー株式会社 | Image processing apparatus, program, image processing method, and terminal device |
KR101706216B1 (en) * | 2012-04-03 | 2017-02-13 | 한화테크윈 주식회사 | Apparatus and method for reconstructing dense three dimension image |
US10130277B2 (en) | 2014-01-28 | 2018-11-20 | Medibotics Llc | Willpower glasses (TM)—a wearable food consumption monitor |
US9456916B2 (en) | 2013-03-12 | 2016-10-04 | Medibotics Llc | Device for selectively reducing absorption of unhealthy food |
US10314492B2 (en) | 2013-05-23 | 2019-06-11 | Medibotics Llc | Wearable spectroscopic sensor to measure food consumption based on interaction between light and the human body |
US20140104385A1 (en) * | 2012-10-16 | 2014-04-17 | Sony Network Entertainment International Llc | Method and apparatus for determining information associated with a food product |
TWI498845B (en) * | 2012-12-14 | 2015-09-01 | Ind Tech Res Inst | Method and system for diet management |
CN103902613A (en) * | 2012-12-30 | 2014-07-02 | 青岛海尔软件有限公司 | Speech recognition and cloud search engine technology based man-machine interactive system and method |
US8690578B1 (en) | 2013-01-03 | 2014-04-08 | Mark E. Nusbaum | Mobile computing weight, diet, nutrition, and exercise tracking system with enhanced feedback and data acquisition functionality |
US9067070B2 (en) | 2013-03-12 | 2015-06-30 | Medibotics Llc | Dysgeusia-inducing neurostimulation for modifying consumption of a selected nutrient type |
US9011365B2 (en) | 2013-03-12 | 2015-04-21 | Medibotics Llc | Adjustable gastrointestinal bifurcation (AGB) for reduced absorption of unhealthy food |
WO2014160298A1 (en) * | 2013-03-14 | 2014-10-02 | Sciencestyle Capital Partners, Llc | Providing food-portion recommendations to faciliate dieting |
JP2014211749A (en) * | 2013-04-18 | 2014-11-13 | ソニー株式会社 | Information processing device and storage medium |
JP6098326B2 (en) * | 2013-04-18 | 2017-03-22 | ソニー株式会社 | Information processing apparatus and storage medium |
US9977988B2 (en) * | 2013-05-03 | 2018-05-22 | Csb-System Ag | Device and method for the classification of a food item of an organic or irregular structure |
US9230194B2 (en) * | 2013-09-16 | 2016-01-05 | Google Inc. | Training image sampling |
US9314206B2 (en) | 2013-11-13 | 2016-04-19 | Memphis Technologies, Inc. | Diet and calories measurements and control |
US9311639B2 (en) | 2014-02-11 | 2016-04-12 | Digimarc Corporation | Methods, apparatus and arrangements for device to device communication |
US9330469B2 (en) | 2014-04-08 | 2016-05-03 | General Electric Company | Systems and methods for boil detection |
US9730647B2 (en) * | 2014-04-15 | 2017-08-15 | Vivint, Inc. | Systems and methods for measuring calorie intake |
US9449220B2 (en) | 2014-04-21 | 2016-09-20 | Haier Us Appliance Solutions, Inc. | Systems and methods for cookware detection |
US9412048B2 (en) | 2014-04-21 | 2016-08-09 | Haier Us Appliance Solutions, Inc. | Systems and methods for cookware detection |
CN105806320B (en) * | 2014-12-29 | 2020-04-21 | 同方威视技术股份有限公司 | Imaging measurement system and imaging measurement method |
US10971031B2 (en) | 2015-03-02 | 2021-04-06 | Fitly Inc. | Apparatus and method for identifying food nutritional values |
US9349297B1 (en) * | 2015-09-09 | 2016-05-24 | Fitly Inc. | System and method for nutrition analysis using food image recognition |
US9146147B1 (en) | 2015-04-13 | 2015-09-29 | Umar Rahim Bakhsh | Dynamic nutrition tracking utensils |
US9424495B1 (en) | 2015-04-16 | 2016-08-23 | Social & Health Research Center | Digital food imaging analysis: system and method to analyze food consumption |
US9824434B2 (en) * | 2015-08-18 | 2017-11-21 | Industrial Technology Research Institute | System and method for object recognition |
DE102016007713A1 (en) | 2016-06-23 | 2017-12-28 | Andreas Franzen | Method and device for determining food intake and eating behavior |
US10579905B2 (en) * | 2017-03-17 | 2020-03-03 | Google Llc | Fully parallel, low complexity approach to solving computer vision problems |
GB2561556A (en) * | 2017-04-13 | 2018-10-24 | Kan Jing | Apparatus for providing information about a food product |
US10832094B2 (en) * | 2018-04-10 | 2020-11-10 | International Business Machines Corporation | Generating hyperspectral image database by machine learning and mapping of color images to hyperspectral domain |
KR102615196B1 (en) | 2018-08-21 | 2023-12-18 | 삼성전자주식회사 | Method and device to train object detection model |
EP3881223A1 (en) * | 2018-11-16 | 2021-09-22 | Snaq GmbH | Device and method for determining and displaying nutrient content and/or value of a food item |
CN111353333B (en) * | 2018-12-21 | 2023-10-20 | 九阳股份有限公司 | Food material identification method, household electrical appliance and food material identification system |
US10909700B2 (en) * | 2019-04-02 | 2021-02-02 | Samsung Electronics Co., Ltd. | Display apparatus and image processing method thereof |
US11410220B2 (en) | 2019-05-15 | 2022-08-09 | Samsung Electronics Co., Ltd. | Exploration for interactive recommendation system, method, and computer program product |
CN110399804A (en) * | 2019-07-01 | 2019-11-01 | 浙江师范大学 | A kind of food inspection recognition methods based on deep learning |
US11733048B2 (en) * | 2019-07-31 | 2023-08-22 | Samsung Electronics Co., Ltd. | Electronic apparatus and control method thereof |
WO2021236884A1 (en) | 2020-05-20 | 2021-11-25 | Passio, Inc. | Food-recognition systems and methods |
US11138410B1 (en) * | 2020-08-25 | 2021-10-05 | Covar Applied Technologies, Inc. | 3-D object detection and classification from imagery |
CN113081014B (en) * | 2021-03-19 | 2023-05-30 | 东软医疗系统股份有限公司 | Spacer scanning method, device and system |
Family Cites Families (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JP2001216133A (en) * | 2000-01-31 | 2001-08-10 | Tanita Corp | Calorie management device for voice recognition |
IL137759A (en) * | 2000-08-08 | 2007-07-16 | Eddie Karnieli | Method for monitoring food intake |
-
2010
- 2010-01-06 US US12/683,124 patent/US8439683B2/en active Active
-
2013
- 2013-03-22 US US13/849,222 patent/US20130260345A1/en not_active Abandoned
-
2017
- 2017-11-20 US US15/818,528 patent/US20180075774A1/en not_active Abandoned
Cited By (20)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20130170714A1 (en) * | 2010-05-31 | 2013-07-04 | The University Of Tokyo | Information processing device |
US9449226B2 (en) * | 2010-05-31 | 2016-09-20 | The University Of Tokyo | Information processing device |
US9042596B2 (en) | 2012-06-14 | 2015-05-26 | Medibotics Llc | Willpower watch (TM)—a wearable food consumption monitor |
US10772559B2 (en) | 2012-06-14 | 2020-09-15 | Medibotics Llc | Wearable food consumption monitor |
US9254099B2 (en) | 2013-05-23 | 2016-02-09 | Medibotics Llc | Smart watch and food-imaging member for monitoring food consumption |
US9529385B2 (en) | 2013-05-23 | 2016-12-27 | Medibotics Llc | Smart watch and human-to-computer interface for monitoring food consumption |
US9536449B2 (en) | 2013-05-23 | 2017-01-03 | Medibotics Llc | Smart watch and food utensil for monitoring food consumption |
US9442100B2 (en) | 2013-12-18 | 2016-09-13 | Medibotics Llc | Caloric intake measuring system using spectroscopic and 3D imaging analysis |
US9977980B2 (en) * | 2014-02-12 | 2018-05-22 | Microsoft Technology Licensing, Llc | Food logging from images |
US11864692B2 (en) | 2014-03-14 | 2024-01-09 | Spectrum Brands, Inc. | Food preparation appliance for use with a remote communication device |
US11272810B2 (en) | 2014-03-14 | 2022-03-15 | Spectrum Brands, Inc. | Food preparation appliance for use with a remote communication device |
US9916520B2 (en) | 2014-09-03 | 2018-03-13 | Sri International | Automated food recognition and nutritional estimation with a personal mobile electronic device |
US9734426B2 (en) | 2014-09-03 | 2017-08-15 | Sri International | Automated food recognition and nutritional estimation with a personal mobile electronic device |
US10521903B2 (en) | 2015-11-25 | 2019-12-31 | Samsung Electronics Co., Ltd. | User terminal apparatus and control method thereof |
US10861153B2 (en) | 2015-11-25 | 2020-12-08 | Samsung Electronics Co., Ltd. | User terminal apparatus and control method thereof |
US11568981B2 (en) | 2015-11-25 | 2023-01-31 | Samsung Electronics Co., Ltd. | User terminal apparatus and control method thereof |
US9972080B2 (en) | 2016-01-14 | 2018-05-15 | Electronics And Telecommunications Research Institute | Apparatus and method for food search service |
US10943398B2 (en) | 2016-07-15 | 2021-03-09 | Samsung Electronics Co., Ltd. | Augmented reality device and operation thereof |
CN107194418A (en) * | 2017-05-10 | 2017-09-22 | 中国科学院合肥物质科学研究院 | A kind of Aphids in Rice Field detection method based on confrontation feature learning |
US11069045B2 (en) | 2018-08-03 | 2021-07-20 | International Business Machines Corporation | Food preparation assistance using image processing |
Also Published As
Publication number | Publication date |
---|---|
US20180075774A1 (en) | 2018-03-15 |
US20100173269A1 (en) | 2010-07-08 |
US8439683B2 (en) | 2013-05-14 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
US20180075774A1 (en) | Food recognition using visual analysis and speech recognition | |
US8345930B2 (en) | Method for computing food volume in a method for analyzing food | |
Puri et al. | Recognition and volume estimation of food intake using a mobile device | |
Bolanos et al. | Simultaneous food localization and recognition | |
Lo et al. | Image-based food classification and volume estimation for dietary assessment: A review | |
Jiang et al. | DeepFood: food image analysis and dietary assessment via deep model | |
Rohrbach et al. | Recognizing fine-grained and composite activities using hand-centric features and script data | |
Meyers et al. | Im2Calories: towards an automated mobile vision food diary | |
Farinella et al. | Retrieval and classification of food images | |
Zhang et al. | “snap-n-eat” food recognition and nutrition estimation on a smartphone | |
Santra et al. | A comprehensive survey on computer vision based approaches for automatic identification of products in retail store | |
Zhu et al. | Multiple hypotheses image segmentation and classification with application to dietary assessment | |
US9734426B2 (en) | Automated food recognition and nutritional estimation with a personal mobile electronic device | |
Zhu et al. | The use of mobile devices in aiding dietary assessment and evaluation | |
US9195904B1 (en) | Method for detecting objects in stereo images | |
Zhu et al. | Segmentation assisted food classification for dietary assessment | |
US11756282B2 (en) | System, method and computer program for guided image capturing of a meal | |
US8094971B2 (en) | Method and system for automatically determining the orientation of a digital image | |
TW201709110A (en) | System and method for object recognition | |
Swadzba et al. | Indoor scene classification using combined 3D and gist features | |
Lee | Component-based face detection and verification | |
Fang et al. | The use of co-occurrence patterns in single image based food portion estimation | |
CN115516519A (en) | Method and apparatus for hierarchy-based object recognition for visual perception | |
Velvizhy et al. | Automatic food recognition system for diabetic patients | |
Swadzba et al. | A detailed analysis of a new 3D spatial feature vector for indoor scene classification |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
AS | Assignment |
Owner name: SARNOFF CORPORATION, NEW JERSEY Free format text: ASSIGNMENT OF ASSIGNORS INTEREST;ASSIGNORS:PURI, MANIKA;ZHU, ZHIWEI;LUBIN, JEFFREY;AND OTHERS;SIGNING DATES FROM 20100201 TO 20100206;REEL/FRAME:030076/0932 Owner name: SRI INTERNATIONAL, CALIFORNIA Free format text: ASSIGNMENT OF ASSIGNORS INTEREST;ASSIGNOR:SARNOFF CORPORATION;REEL/FRAME:030077/0277 Effective date: 20110204 |
|
STCB | Information on status: application discontinuation |
Free format text: ABANDONED -- AFTER EXAMINER'S ANSWER OR BOARD OF APPEALS DECISION |