Research On Image Generation Of Dairy Goat Based On Improved-SNGAN

Deep learning needs a large number of data samples when training the network.Generative Adversarial Nets(GAN)is a deep learning model.Image generation and data enhancement are the main applications of GAN.Aiming at the problem of serious shortage of training image samples for the existing dairy goats,this article takes the image of the dairy goat farm in the Animal Husbandry Teaching and Experiment Base of Northwest A&F University as the research object,uses the spectrum constraint normalization to generate the confrontation network model,and combines the residual module to realize the milk The generation of goat image samples to expand the dairy goat data set.The main research contents and conclusions of this article are:(1)Collection and preprocessing of dairy goat dataFirst,the data of dairy goats are obtained by video interception and manual photographing.Then artificially screen the clear image of the dairy goat on the side and the whole body to eliminate the information irrelevant to the experiment,and classify it according to the posture and background.Finally,the simple data enhancement of the dairy goat image by flipping and cropping is carried out to expand the dairy goat data set used in this experiment.(2)Research on the generation of dairy goat image based on SNGANThis article proposes a method for generating samples of dairy goat images based on SNGAN.By limiting the spectral norm of each layer of the network,SNGAN enables the generation model to be more fully trained and achieves the effect of balancing against the discriminative model,thereby enhancing the training of GAN Stability in the process,reducing mode crashes.At the same time,an auxiliary classifier is introduced to add label information to the original GAN to enrich the diversity of the generated images,and to improve the sample quality through small-batch discrimination.(3)Research on the generation of dairy goat image samples based on the residual blocksA method for generating samples of dairy goat images based on the residual structure is proposed.The residual structure is added to the generation model and the discriminant model to deepen the network depth,extract deeper feature information,and strengthen information transmission.And redesigned two residual structures to make them more suitable for the study of this article.Finally,an experimental comparison is made with the original GAN and three GAN derived models.The Inception Score(IS)and Fréchet Inception Distance(FID)of this method are better than the other four models.The images of dairy goats are also clearer and more diverse.In summary,this article has conducted an in-depth study on the generation of dairy goat images for the generation of confrontation,and proposed improved methods for the instability of model training and model collapse.Experiments have proved the effectiveness of the method in this article.The scale of the dairy goat image data.