GCT-VAE-GAN: An Image Enhancement Network for Low-Light Cattle Farm Scenes by Integrating Fusion Gate Transformation Mechanism and Variational Autoencoder GAN

In the context of cattle farm environments, intricate environmental interferences have presented challenges that impede seamless data acquisition. This paper introduces a novel approach, the integration of a fusion gate transformation mechanism and a variational autoencoder GAN, which we term GCT-VAE-GAN, aimed at enhancing low-light images from cattle farm settings. Initially, our approach involves the design of an encoding network tasked with augmenting the original images. Subsequently, we advance our methodology by formulating a generative network to effectively address the challenges of image diversification and poor image quality. Notably, the inclusion of an attention mechanism block within the FFN layer facilitates the fusion of these extracted features, resulting in the generation of high-quality images. Furthermore, to achieve proficient image discrimination, we implement a dual-discriminator structured discriminative network for the conclusive image discrimination task. The culmination of our approach involves the formulation of a comprehensive joint loss function, thereby constituting the core of the model's loss module. Moreover, through comparative experiments, we aptly demonstrate the remarkable superiority of the GCT-VAE-GAN approach. The conducted experiments reveal the model's consistent performance and resilience under varying illumination scenarios. The outcomes of our study underscore its significant relevance in elevating the quality of low-light images within cattle farm contexts. Furthermore, our approach exhibits the potential to enhance the efficacy of computer vision tasks, signifying a notable stride toward improved agricultural imaging techniques.