Validating Time Serial Images for Emotion Recognition

The pig emotion recognition (PER) reads the pig's emotions through the surveillance camera system. It can notify the husbandry workers if the system finds the pig's negative emotions. The PER is the idealistic system for pig husbandry workers, but practically applying the system is quite challenging without a suitable PER dataset. The pigs in the cage seldom move around the area, and the images captured are almost identical through the surveillance camera in the time series. The recurrent convolution neural networks may solve the problem. Still, with the single inputs to the architecture, the time serial images of the PER dataset produce misleading and biased experimental narration without a proper preprocessing method. To have adequate results from the time-serial imaging dataset, we propose the semi-shuffling approach to manage our PER dataset rather than what some researchers normally fully-shuffle the dataset without inspecting time-serial images. We have 98.45% validating accuracy as with fully-shuffling whole training and testing groups, but the validating accuracy reduces to 75.97% after applying the semi-shuffling training and testing dataset.