影像去背(Image Matting)是許多影像應用的基礎，如電影合成、遠距會議…等。影像去背可以分成三種去背問題，單色去背、差異去背、自然影像去背。
單色去背的影像背景為已知的單一顏色，通常用於拍攝電影時，藉由綠幕的方式替換背景，為最簡單的影像去背。差異去背需要拍攝完已知背景後，再拍攝加入人物後的影像，此方法在邊界上需要加強處理。自然影像去背的背景沒有限制，同時也是最困難的去背問題，此問題主要是希望能找出影像中欲保留的前景，因為一張影像可能同時包含了前景、背景、非預期的前景，但只要保留指定的前景，因此需要找到一張前景遮照，藉由此遮罩影像來獲得欲保留的前景。早期的方法需要藉由人工手動標記前景與背景，這樣的方法非常消耗時間。
近年的方法也在輸入影像有著許多條件限制，在應用上並不方便。本篇論文結合許多影像演算法和神經網路的方式，僅需給予一張照片或影片，並且在一開始手動框選欲保留的人物，將輸入影像經過一系列運算後得到該遮罩影像，藉由該遮罩影像與原始影像結合得到前景。
在本篇論文先藉 DeepLabV3 [1]得到人的粗糙前景遮罩，該影像在邊界處理是非常粗糙的，因此透過新訓練的模型來改善邊界的粗糙問題。DeepLabV3 [1]會同時找出所有人的粗糙前景遮罩，為了抓取單一人物的粗糙前景遮罩，採取物件追蹤[2]、物件連通[3]與影像分割[4]的技術，獲得單一人物的粗糙前景遮罩。
實驗結果發現，由本篇論文所提出的方法克服了其他影像去背方法所遇到的問題，不會因為動態背景或是聚焦問題而影響結果，也不需要做手動標記前景背景，僅需在一開始框選欲追蹤的人物。由於問題複雜度上升，整體計算量也隨之上升，但克服了許多方法的限制條件。

Image Matting is the basis of many image applications, such as film compositing,
teleconferencing. There are three types of problems in image matting: difference
matting, natural image matting, and constant-color matting. Earlier methods required
manual labeling of foreground and background, which was time-consuming. The
methods in recent years also have many constraints on the input image, which are
inconvenient in application. This thesis combines many image algorithms and neural
network methods to solve natural image matting. It only needs to give a photo or video,
and manually select the characters to be retained at the beginning. After a series of the operations, the mask is obtained. We can combine the mask and original image to obtain
the foreground.
In our method, DeepLabV3 [1] creates the rough foreground mask which is very
rough at the boundary. We train a new model to improve the rough boundary.
DeepLabV3 [1] will find all people in mask. In order to capture specified person’s mask, we use object tracking [2], connected-component labeling [3] and image segmentation
[4] to obtain the specified one person.
Our experimental results show that our method overcomes the problems from
other image matting methods, and does not affect the results due to dynamic
background or focus problems, and does not require manual marking of the foreground
and background. It only needs frame the person you want to track. As the complexity
of the problem increases, such as rough edges, unexpected person, and overlapping
person. The overall computational load increase, but it overcomes the limitations of
many methods.

[1] Chen, L. C., Papandreou, G., Schroff, F., & Adam, H. (2017). Rethinking atrous
convolution for semantic image segmentation. (arXiv preprint arXiv: 1706.05587).
[2] Bolme, D. S., Beveridge, J. R., Draper, B. A., & Lui, Y. M. (2010). Visual object
tracking using adaptive correlation filters. IEEE Computer Society Conference on
Computer Vision and Pattern Recognition (pp. 2544-2550).
[3] He, L., Chao, Y., Suzuki, K., & Wu, K. (2009). Fast connected-component labeling.
Pattern Recognition, 42(9), (pp. 1977-1987).
[4] Achanta, R., Shaji, A., Smith, K., Lucchi, A., Fua, P., & Süsstrunk, S. (2012).
SLIC Superpixels Compared to state-of-the-art superpixel methods. IEEE
Transactions on Pattern Analysis and Machine Intelligence, (pp. 2274-2282).
[5] Chuang, Y. Y., Curless, B., Salesin, D. H., & Szeliski, R. (2001). A bayesian
approach to digital matting. IEEE Computer Society Conference on Computer
Vision and Pattern Recognition. CVPR 2001. (Vol. 2, pp. II-264).
[6] Sun, J., Jia, J., Tang, C. K., & Shum, H. Y. (2004). Poisson matting. ACM
SIGGRAPH (pp. 315-321).
[7] Zheng, Y., & Kambhamettu, C. (2009). Learning based digital matting. IEEE 12th
International Conference on Computer Vision (pp. 889-896).
[8] Singh, S., Jalal, A. S., & Bhatanagar, C. (2013). Automatic trimap and alpha-matte
generation for digital image matting. Sixth International Conference on
Contemporary Computing (IC3) (pp. 202-208).
[9] Canny, J. (1986). A computational approach to edge detection. IEEE Transactions
on Pattern Analysis and Machine Intelligence (pp. 679-698).
[10] Xu, N., Price, B., Cohen, S., & Huang, T. (2017). Deep image matting. IEEE
Conference on Computer Vision and Pattern Recognition (pp. 2970-2979).
[11] Sengupta, S., Jayaram, V., Curless, B., Seitz, S. M., & Kemelmacher-Shlizerman,
I. (2020). Background matting: The world is your green screen. IEEE/CVF
Conference on Computer Vision and Pattern Recognition (pp. 2291-2300).
[12] Lin, S., Ryabtsev, A., Sengupta, S., Curless, B. L., Seitz, S. M., & KemelmacherShlizerman, I. (2021). Real-time high-resolution background matting. IEEE/CVF
Conference on Computer Vision and Pattern Recognition (pp. 8762-8771).
[13] He, K., Zhang, X., Ren, S., & Sun, J. (2016). Deep residual learning for image
recognition. IEEE Conference on Computer Vision and Pattern Recognition (pp.
770-778).
[14] Chen, L. C., Papandreou, G., Kokkinos, I., Murphy, K., & Yuille, A. L. (2017).
Deeplab: Semantic image segmentation with deep convolutional nets, atrous
convolution, and fully connected crfs. IEEE Transactions on Pattern Analysis and
Machine Intelligence, 40(4), (pp. 834-848).
[15] Li, J., Zhang, J., Maybank, S. J., & Tao, D. (2020). End-to-end animal image
matting. (arXiv preprint arXiv:2010.16188).
[16] Chang, H. T., & Su, J. T. (2016). Segmentation using Codebook Index Statistics
for Vector Quantized Images. International Journal of Advanced Computer
Science and Applications , 7(12), (pp. 59-65).
[17] Lin, T. Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., et al. (2014).
Microsoft coco: Common objects in context. In 2014 European Conference on
Computer Vision (pp. 740-755).