Mouse Behavior Recognition Based on Convolution Neural Network

Mouse behavior recognition has an important role in basic and clinical neuroscience. The videos recording the animals are often annotated manually. The method of traditional manual annotation behavior is time-consuming, laborious, subjective and poorly reproducible. If the traditional 2D convolutional neural network is used to recognize mouse behavior, it will be easy to lose time dimension information. At present, 3D CNN method has been proposed for action recognition. However, the method has the disadvantages of high computational cost and large training samples demand. This paper summarizes videos into single images and develops a 2D CNN model to analyze mouse behavior by using these images as model inputs. Therefore, the behavior recognition model can extract features from both spatial and temporal dimensions. It is trained by fine-tuning the CaffeNet. This method is simple but powerful because of utilizing 2D CNN models directly on video data. Through experiments, this paper demonstrates the model can provide more accurate annotations than other automatic methods.