將殘差塊應用於生成對抗網路訓練及穩定性之研究

近年來人工智慧迅速發展，許多行業紛紛引進相關技術，例如語音辨識、汽車的自動駕駛等等，生成對抗網路在無監督學習任務中表現出令人印象深刻的性能，是現在最熱門的生成模型，它將生成建模視為兩人遊戲，生成器神經網路在給定一些隨機噪聲的情況下學習，以生成近似於真實樣本的合成樣本，而判別器的神經網路區分生成器的輸出或是真實樣本，到目前為止，生成對抗網路有著廣泛的應用，已有大量研究表明可以在各種任務中發揮重要的作用，自然語言處理可應用在文本生成圖像、機器翻譯、對話生成，而研究最多的是圖像處理領域，例如圖像風格遷移、圖像超分辨率、圖像修復等。儘管生成對抗網路在圖像生成方面獲得顯著的成功，卻仍存在著訓練不穩定和模型崩潰，為了穩定生成對抗網路的訓練過程並提高生成圖像的多樣性，本文研究對生成器與判別器的神經網路進行改動，將殘差塊加入生成對抗網路架構學習圖像特徵，減少訓練過程中丟失的圖像特徵訊息，獲得更多的特徵以穩定圖像生成，以特徵匹配的方法最小化生成圖像跟真實圖像之間的差異，加入梯度約束以減輕訓練過程中梯度消失的問題。

關鍵字

圖像生成；生成對抗網路；殘差塊

並列摘要

In the last few years, artificial intelligence has developed rapidly and many industries have introduced related technologies, such as speech recognition and autonomous cars and so on. Generative adversarial networks have demonstrated impressive performance for unsupervised learning tasks and are now the most popular generative model, which treats generative model as two-player game, generator neural network learns given some random noise to generative synthetic samples similar to real samples, and discriminator neural network discriminates between the generator output and real samples. So far, plenty of works have shown that generative adversarial networks can play a significant role in various tasks, including natural language processing for text to image generation, machine translation, dialogue generation, and the most research is image process, such as image style transfer, image super-resolution and image restoration and so on. Although generative adversarial networks remarkable success in image generation, but training process is usually unstable and model collapse, in order to stabilize the generative adversarial networks, training process and improve diversity of generated images, In this paper, the neural network of generator and discriminator are modified, and the residual block is added to the generative adversarial network architecture to learn image features. Reduce the loss of image feature during training, and get more features to stabilize image generation, feature matching is used to minimize the difference between the generated images and real images, and the gradient constraint is added to mitigate the problem of gradient vanishing during training.