Research On Semantic Segmentation Of Dairy Goat Video Based On Fully Convolutional Networks

Semantic segmentation of the monitoring video of the outdoor dairy goat house scene of the dairy goat,observing the location and posture of the sheep house environment and the dairy goat,can detect the health problems of the dairy goat in time,and find out whether the sheep house has the invasion of foreign matter that is safe for the sheep.The large-scale and intelligent breeding of dairy goats is of great significance.This paper takes the dairy goat surveillance video as the research object,and uses the video key frame extraction technology,Fully Convolutional Network（FCN）,and the Generated Confrontation Network（GAN）to realize the video of the dairy goat.Semantic segmentation.The main research contents and conclusions of this paper are as follows:（1）Production of dairy goat video semantic segmentation model data setIn order to reduce the large amount of redundant information existing between video frames,the Chi-square（?²）histogram method combined with the adaptive threshold method is firstly used to realize the lens segmentation of the surveillance video of the dairy goat.The keyframes are then extracted from each shot using the k-means clustering method to create an image dataset for semantic segmentation.In addition,in order to increase the data set,this paper uses the flip transform and translation transform method to enhance the data,and expand the data set to 5 times,thus preventing the over-fitting of semantic segmentation results.（2）Design of video semantic segmentation model for dairy goats based on FCNFor the problem of speed and precision of convolutional neural networks for semantic segmentation,VGG₁6 is used as a pre-training network to perform semantic segmentation of dairy goat video using full convolutional neural network,extending from image level classification to pixel level.Classification.In addition,the segmentation results of FCN still have shortcomings such as insufficiency and lack of spatial consistency.The coarse segmentation result of FCN₈s model is segmented by Conditional Random Field（CRF）to obtain a relationship between pixels.With more detailed semantic segmentation results,the pixel accuracy of semantic segmentation is 0.85%higher than FCN₈s,and the average cross-ratio is increased by 0.6%.（3）Design of video semantic segmentation model of dairy goat based on FDGANAiming at FCN₈s+CRF is a network model of coarse segmentation and fine segmentation,and the semantic segmentation results still have leakage and roughness.This paper proposes FDGAN network model.This model adopts the idea of GAN,uses FCN₈s as the generation network of GAN,generates semantic segmentation results,and uses Dense Convolutional Network（DenseNet）as the discriminant network.Through the confrontation training of two networks,semantically segmented pixels Accuracy is 2.57%higher than FCN₈s,and the average cross-over ratio is increased by 2.36%.FDGAN has obtained more detailed and spatially consistent segmentation results.