使用小波轉換卷積神經網絡進行RGB-D影像深度訊息的可逆嵌入

Abstract

隨著各種深度攝影機與三維虛擬實境顯示裝置的普及，帶有深度資訊的三維 影像也越來越廣泛地被使用。但是這種新型資料格式尚未有統一傳輸存儲格式， 並且在傳統設備上也存在相容性問題。 因此本論文提出了，一種全新的將三維圖片的深度資訊，嵌入在二維通道內 並可解碼復原的方法。本文方法中的神經網絡骨幹，為卷積神經網路與離散小波 轉換之結合。 相較於傳統單獨使用卷積神經網路的編碼器與解碼器網路，有著佔用記憶體 空間小，運算速度快，編解碼品質佳，這些優點。除此之外，對於可逆圖像轉換 問題的端到端訓練中，遇到的量化誤差，本文提出了一種全新的可求導的量化函 數，使得神經網路在測試時更少地受到量化誤差的影響。

With the popularization of various depth cameras and 3D virtual reality monitors, 3D images with depth information are more and more widely used. However, there is no such unified format for these new types of data for transmission and storage. Moreover, using traditional devices may cause compatibility issues. Therefore, in this thesis, we propose a novel method of embedding the depth information of 3D images into RGB channels, which can be restored back. The backbone of the neural network in this method is a combination of a convolutional neural network and a discrete wavelet transform module. Compared with traditional encoder and decoder networks purely using convolution neural networks alone, the advantages of our method are less memory demanding, high computing speed, and better visual quality for encoding and decoding. Additionally, to solve the quantization error in the end-to-end training of the reversible image conversion problems, we propose a differentiable quantization function. So that the neural network is less influenced by quantization error in the testing stage.