Download files
Abstract
Recently obesity has become an epidemic and one of the most serious worldwide public
health concerns of the 21st century. Obesity diminishes the average life expectancy and
there is now convincing evidence that poor diet, in combination with physical inactivity
are key determinants of an individual s risk of developing chronic diseases such as
cancer, cardiovascular disease or diabetes. Assessing what people eat is fundamental
to establishing the link between diet and disease. Food records are considered the best
approach for assessing energy intake. However, this method requires literate and highly
motivated subjects. This is a particular problem for adolescents and young adults who
are the least likely to undertake food records. The ready access of the majority of the
population to mobile phones (with integrated camera, improved memory capacity, network
connectivity and faster processing capability) has opened up new opportunities for
dietary assessment. The dietary information extracted from dietary assessment provide
valuable insights into the cause of diseases that greatly helps practicing dietitians and
researchers to develop subsequent approaches for mounting intervention programs for
prevention. In such systems, the camera in the mobile phone is used for capturing images
of food consumed and these images are then processed to automatically estimate the nutritional content of the food. However, food objects are deformable objects that
exhibit variations in appearance, shape, texture and color so the food classification and
volume estimation in these systems suffer from lower accuracy. The improvement of
the food recognition accuracy and volume estimation accuracy are challenging tasks.
This thesis presents new techniques for food classification and food volume estimation.
For food recognition, emphasis was given to texture features. The existing food
recognition techniques assume that the food images will be viewed at similar scales and
from the same viewpoints. However, this assumption fails in practical applications, because
it is difficult to ensure that a user in a dietary assessment system will put his/her
camera at the same scale and orientation to capture food images as that of the target food
images in the database. A new scale and rotation invariant feature generation approach
that applies Gabor filter banks is proposed. To obtain scale and rotation invariance,
the proposed approach identifies the dominant orientation of the filtered coefficient and
applies a circular shifting operation to place this value at the first scale of dominant
direction. The advantages of this technique are it does not require the scale factor to
be known in advance and it is scale/and rotation invariant separately and concurrently.
This approach is modified to achieve improved accuracy by applying a Gaussian window
along the scale dimension which reduces the impact of high and low frequencies of
the filter outputs enabling better matching between the same classes. Besides automatic
classification, semi automatic classification and group classification are also considered
to have an idea about the improvement. To estimate the volume of a food item, a stereo pair is used to recover the structure as a 3D point cloud. A slice based volume estimation
approach is proposed that converts the 3D point cloud to a series of 2D slices.
The proposed approach eliminates the problem of knowing the distance between two
cameras with the help of disparities and depth information from a fiducial marker. The
experimental results show that the proposed approach can provide an accurate estimate
of food volume.