Brightness Adaptive Food Recognition Using CNN

With the advance of technology, people's quality of life is getting better and better, but the number of death due to illness gradually increased. At the top ten causes of death, diabetes and kidney diseases account for the fifth and ninth respectively. The two diseases in the following treatment are strict in the diet control. At present, although there are many wearable devices able to calculate how much calories consumed, the wearing device estimating the calories and sodium content of food for the user is not yet popular since food recognition is still challenging due to the variety of their color, shape, and texture. Among them, brightness is one of the most critical problems. In view of this, we employ the brightness to calculate the reflective region, and adjust its hue and saturation without changing the texture. Hence, the region can be consistent with other nonreflective region of color consistency. In the food recognition, we use convolution neural network to extract the simple to complex features, and convert it to fisher vector through fisher kernel. Eventually, we utilize support vector machine for classification.