Compact neural network design for visual food recognition

Deep learning grows fast in recent years, and people already use it to solve many real-world problems. Convolutional neural network (CNN) is one of the essential architectures of deep learning. Compared to traditional machine learning methods, CNN can generate transformed feature maps automatically. With this, the machine improves its self-learning ability remarkably. In this thesis, we utilize CNN to perform visual food recognition. After that, our result could serve as a backbone to health monitoring smartphone apps, which can recognize a variety of foods and show their nutritional components. The problem encountered is that deep CNNs such as VGG and ResNet are too large to implement on low-end devices (e.g., smartphones). Hence, we try to find one compact architecture that has competitive accuracy with a little computational cost. We select several state-of-the-art compact CNNs such as MobileNet, ShuffleNet, and SqueezeNet. We mainly compare their performances with a deep neural network — ResNet50, including accuracy, the number of parameters, and the floating-point operations per second (FLOPS). Finally, we choose the most suitable architecture for visual food recognition and tune it to improve the performance further.Master of Science (Signal Processing