Datasets, features, learning, and models in visual recognition

Visual recognition is a fundamental research topic in computer vision. This dissertation explores datasets, features, learning, and models used for visual recognition. In order to train visual models and evaluate different recognition algorithms, this dissertation develops an approach to collect object image datasets on web pages using an analysis of text around the image and of image appearance. This method exploits established online knowledge resources (Wikipedia pages for text; Flickr and Caltech data sets for images). The resources provide rich text and object appearance information. This dissertation describes results on two datasets. The first is Berg’s collection of 10 animal categories; on this dataset, we significantly outperform previous approaches. On an additional set of 5 categories, experimental results show the effectiveness of the method. Images are represented as features for visual recognition. This dissertation introduces a text-based image feature and demonstrates that it consistently improves performance on hard object classification problems. The feature is built using an auxiliary dataset of images annotated with tags, downloaded from the Internet. Image tags are noisy. The method obtains the text features of an unannotated image from the tags of its k-nearest neighbors in this auxiliary collection. A visual classifier presented with an object viewed under novel circumstances (say, a new viewing direction) must rely on its visual examples. This text feature may not change, because the auxiliary dataset likely contains a similar picture. While the tags associated with images are noisy, they are more stable when appearance changes. The performance of this feature is tested using PASCAL VOC 2006 and 2007 datasets. This feature performs well; it consistently improves the performance of visual object classifiers, and is particularly effective when the training dataset is small. With more and more collected training data, computational cost becomes a bottleneck, especially when training sophisticated classifiers such as kernelized SVM. This dissertation proposes a fast training algorithm called Stochastic Intersection Kernel Machine (SIKMA). This proposed training method will be useful for many vision problems, as it can produce a kernel classifier that is more accurate than a linear classifier, and can be trained on tens of thousands of examples in two minutes. It processes training examples one by one in a sequence, so memory cost is no longer the bottleneck to process large scale datasets. This dissertation applies this approach to train classifiers of Flickr groups with many group training examples. The resulting Flickr group prediction scores can be used to measure image similarity between two images. Experimental results on the Corel dataset and a PASCAL VOC dataset show the learned Flickr features perform better on image matching, retrieval, and classification than conventional visual features. Visual models are usually trained to best separate positive and negative training examples. However, when recognizing a large number of object categories, there may not be enough training examples for most objects, due to the intrinsic long-tailed distribution of objects in the real world. This dissertation proposes an approach to use comparative object similarity. The key insight is that, given a set of object categories which are similar and a set of categories which are dissimilar, a good object model should respond more strongly to examples from similar categories than to examples from dissimilar categories. This dissertation develops a regularized kernel machine algorithm to use this category dependent similarity regularization. Experiments on hundreds of categories show that our method can make significant improvement for categories with few or even no positive examples.

In vivo and in vitro disappearance of energy and nutrients in novel carbohydrates and cereal grains by pigs

In vivo and in vitro experiments were conducted to determine digestibility of GE and nutrients, as well as DE and ME of carbohydrates fed to growing pigs. The objective of Exp. 1 was to determine the DE and ME of 4 novel carbohydrates fed to pigs. The 4 novel carbohydrates were 2 sources of resistant starch (RS 60 and RS 70), soluble corn fiber (SCF), and pullulan. These carbohydrates were produced to increase total dietary fiber (TDF) intake by humans. Maltodextrin (MD) was used as a highly digestible control carbohydrate. The DE and ME for RS 60 (1,779 and 1,903 kcal/kg, respectively), RS 75(1,784 and 1,677 kcal/kg, respectively), and SCF (1,936 and 1,712 kcal/kg, respectively) were less (P < 0.05) than for MD (3,465 and 3,344 kcal/kg, respectively) and pullulan (2,755 and 2,766 kcal/kg, respectively), and pullulan contained less (P < 0.05) DE and ME than MD. However, there was no difference in the DE and ME for RS 60, RS 75, and SCF. The varying degrees of small intestinal digestibility and differences in fermentability among these novel carbohydrates may explain the differences in the DE and ME among carbohydrates. Therefore, the objectives of Exp. 2 were to determine the effect of these 4 novel carbohydrates and cellulose on apparent ileal (AID) and apparent total tract (ATTD) disappearance, and hindgut disappearance (HGD) of GE, TDF, and nutrients when added to diets fed to ileal-cannulated pigs. The second objective was to measure the endogenous flow of TDF to be able to calculate the standardized ileal disappearance (SID) and standardized total tract (STTD) disappearance of TDF in the 4 novel fibers fed to pigs. Results of the experiment indicated that the AID of GE and DM in diets containing cellulose or the novel fibers was less (P < 0.05) than of the maltodextrin diet, but the ATTD of GE and DM was not different among diets. The addition of RS 60, RS 75, and SCF did not affect the AID of acid hydrolysed ether extract (AEE), CP, or ash, but the addition of cellulose and pullulan reduced (P < 0.01) the AID of CP. The average ileal and total tract endogenous losses of TDF were calculated to be 25.25 and 42.87 g/kg DMI, respectively. The SID of TDF in diets containing RS 60, SCF, and pullulan were greater (P < 0.01) than the SID of TDF in the cellulose diet, but the STTD of the SCF diet was greater (P < 0.05) than for the cellulose and pullulan diets. Results of this experiment indicate that the presence of TDF reduces small intestinal disappearance of total carbohydrates and energy which may reduce the DE and ME of diets and ingredients. Therefore, the objective of Exp. 3 was to determine the DE and ME in yellow dent corn, Nutridense corn, dehulled barley, dehulled oats, polished rice, rye, sorghum, and wheat fed to growing pigs and to determine the AID and ATTD of GE, OM, CP, AEE, starch, total carbohydrates, and TDF in these cereal grains fed to pigs. Results indicated that the AID of GE, OM, and total carbohydrates was greater (P < 0.001) in rice than in all other cereal grains. The AID of starch was also greater (P < 0.001) in rice than in yellow dent corn, dehulled barley, rye, and wheat. The ATTD of GE was greater (P < 0.001) in rice than in yellow dent corn, rye, sorghum, and wheat. With a few exceptions, the AID and ATTD of GE and nutrients in Nutridense corn was not different from the values for dehulled oats. Likewise, with a few exceptions, the AID, ATTD, and HGD of GE, OM, total carbohydrates, and TDF in yellow corn, sorghum, and wheat were not different from each other. The AID of GE and AEE in dehulled barley was greater (P < 0.001) than in rye. The ATTD of GE and most nutrients was greater (P < 0.001) in dehulled barley than in rye. Dehulled oats had the greatest (P < 0.001) ME (kcal/kg DM) whereas rye had the least ME (kcal/kg DM) among the cereal grains. Results of the experiment indicate that the presence of TDF and RS may reduce small intestinal digestibility of starch in cereal grains resulting in reduced DE and ME in these grains. Digestibility experiments involving animals are time consuming and expensive. Therefore, the objective of Exp. 4 was to correlate DM and OM digestibility obtained from 3 in vitro procedures with ATTD of GE and with the concentration of DE in 50 corn samples that were fed to growing pigs. The second objective was to develop a regression model that can predict the ATTD of GE or the concentration of DE in corn. The third objective was to evaluate the suitability of using the DaisyII incubator as an alternative to the traditional water bath when determining in vitro DM and OM digestibility. Results indicated that corn samples incubated with Viscozyme for 48 h in the DaisyII incubator improved (P < 0.001) the ability of the procedure to detect small differences in the ATTD of GE or to detect small differences in the concentration of DE in corn. Likewise, compared with using cellulase or fecal inoculum, the variability in the ATTD of GE and the variability in the DE in corn was better (R2 = 0.56; P < 0.05 and R2 = 0.53; P < 0.06, respectively) explained if Viscozyme was used than if cellulase or fecal inoculum was used. A validated regression model that predicted the DE in corn was developed using Viscozyme and with the corn samples incubated in the DaisyII incubator for a 48 h. In conclusion, this present work used the pig as a model for human gastrointestinal function and evaluates carbohydrates from 2 different nutritional perspectives – humans and animals. The addition of novel carbohydrates reduced the digestibility of energy in the diets without necessarily reducing the digestibility of other nutrients. Thus, supplementation of novel carbohydrates in the diets may be beneficial for the management of diabetes. Aside from diabetic management, cereal grains such as rye and sorghum, may also help in BW management because of there low caloric value, but for undernourished individuals, dehulled oats, dehulled barley, and rice are the ideal grains. From an animal nutrition standpoint, high concentration of dietary fiber is undesirable because it reduces feed efficiency. Therefore, the inclusion of feed ingredients that have a high concentration of dietary fiber is often limited in animal diets. Although in vivo determination is ideal, in vitro procedures are useful tools to determine caloric value of food and feed ingredients.