| Summary: | 碩士 === 國立高雄科技大學 === 電機工程系 === 107 === When we want to obtain training data for deep learning, it is very high possible to meet the problems of less training data. As mentioned in some literatures, when they are sampling samples of sound waves, because of the lack of samples, they cause under-fitting during training. In order to solve this phenomenon, the “Data Argumentation” is applied to the data preprocessing, thereby simulating different angles, light rays and object types seen by computer at different distance under the same object.
Recent years, neural network has evolved to a mature state, such as GoogleNet, VGG16 etc. In this thesis, AlexNet is engaged as the main neural networks architecture for real-time object detection. The advantage of this neural network is to have very small structure, and it is expected to have faster data training speed. However, AlexNet can’t mark objects for detecting. We join the regional proposal network method to mark the objects, so that we can more easily know the situation and accuracy of the object detection through the feedback images.
In this thesis, we will introduce how to improve the under-fitting method, and introduce the normalization method and the loss function selection in the neural network one by one. The influence of different functions and parameters in the neural network is analyzed through visualize methods. Visual tools make it easier to adjust the neural network for the cost reduction of consumption time at the simulation of different parameters.
|
|---|
