基於交叉關聯與注意力模組之物件追蹤與嵌入式硬體之卷積層量化;Cross Correlation and Attention Module Based Object Tracking and Convolutional Layer Quantization for Embedded Hardware

NCUIR > College of Electrical Engineering & Computer Science > Graduate Institute of Computer Science and Information Engineering > Electronic Thesis & Dissertation > Item 987654321/86762

Please use this identifier to cite or link to this item: http://ir.lib.ncu.edu.tw/handle/987654321/86762

Title:	基於交叉關聯與注意力模組之物件追蹤與嵌入式硬體之卷積層量化;Cross Correlation and Attention Module Based Object Tracking and Convolutional Layer Quantization for Embedded Hardware
Authors:	李文郁;Li, Wen-Yu
Contributors:	資訊工程學系
Keywords:	物件偵測;物件追蹤;嵌入式硬體;網路量化;Object Detection;Object Tracking;Embedded Hardware;Network Quantization
Date:	2021-08-26
Issue Date:	2021-12-07 13:11:29 (UTC+8)
Publisher:	國立中央大學
Abstract:	深度學習的發展與高運算的支持，對於影像處理的領域產生了各種不同的任務，使得許多的應用場合也陸續提出。其中，物件追蹤的任務是熱門的議題，也是具有挑戰性的任務之一。透過結合物件偵測的模組與追蹤演算法，在連續輸入的圖像中選取物件位置並標注其物件身份，達到物件追蹤的主要目的。此外，深度學習的大量運算行為，模型的部署也備受關注，為了順利將影像的專案可運行於邊緣裝置，模型的整數量化研究與嵌入式硬體的相互整合，儼然也成為了深度學習應用的主要課題。本篇分別探討物件追蹤與影像分類的量化。以追蹤任務而言，採用一階段系統為主，結合物件偵測與追蹤嵌入分支的多任務訓練，不僅提升訓練與推論速度，模型也較兩階段的組織更為簡易且不失精準性。除了一階段追蹤架構之外，也加入交叉關聯平衡與注意力模組，並適度地調整模組與網路結構，試圖克服端對端網路多任務訓練問題與強化物件追蹤的嵌入身份分支。影像分類的整數量化方面，結合了自定義的FPGA卷積單元與量化後的物件分類模型，佐證模型實現於嵌入式硬體的可能性。本篇追蹤實驗結果以不同的網路組合結構、調整嵌入身份維度與更換不同主幹網路，以評估物件追蹤系統的性能與身份識別的差異性。結果顯示，交叉關聯平衡與注意力的組合之下，驗證物件追蹤身份識別的提升。在不同的身份嵌入的維度設定中，不僅能減少運算量，也能提升追蹤效果。而在更換主幹網路後的實驗中，DLA-34與ResNet-101的主幹網路相比之下，驗證DLA-34同時具有少量參數與較好的準確率。本篇的硬體量化實驗中，在量化後的卷積功能與FPGA整合後，得到了以VGG網路架構推論物件分類的結果，瞭解到量化對於影像推論的重要性。 ;The developments of deep learning and high computing support have produced a variety of tasks in the image processing area, and lots of applications have been proposed one after another. In these tasks, object tracking is a hot topic, and it’s also one of the challenging tasks. Through a combination of object detection and tracking algorithm modules, the position of the object is selected in the continuous input image, and the object identity is marked to achieve the purpose of the object tracking problem. In addition, a large number of computing behaviors for deep learning and the development of models have also attracted attention. To smoothly deploy the image project to the edge device, integrating integer quantization research for model and embedded hardware seems to become the main topic of deep learning applications. This article discusses object tracking and the quantization of image classification. In terms of tracking tasks, we adopt a one-stage system, which is combined with the object tracking method and the embedding branch of multi-object tracking. This structure not only improves the speed of training and inference, but the model is also simpler and more accurate than the second stage architecture. In addition, this object detection structure is adopted using the center point to describe the object location, which effectively solves the accuracy of object detection than previous bounding box solution. In addition to the one-stage object tracking architecture, this paper also adds extra cross-correlation balance and attention modules, moderately adjust modules and network structure to attempt to get over multi-task training problem and strengthen the branch of embedding identity for object tracking. In the aspect of image integer quantization, we combine the custom FPGA convolution unit and the quantized object classification model to prove the possibility of implementation in embedded hardware. The results of these tracking experiments evaluate the difference between the performance of object tracking system and identity recognition by using different network combination structures, adjusting the embedded identity dimension, and replacing different backbone networks. The results show that under the combination of cross-correlation balance and attention mechanism, we verified that object tracking identity has been improved. Additionally, it also shows different identity dimension settings, which can’t only reduce the amount of computation but improve the tracking effect. At last, in the experiment of replacing the backbone network, we compare DLA-34 with ResNet-101 backbone network, which shows that DLA-34 has fewer parameters and better accuracy. In the hardware quantization experiments, the results of inferring object classification based on different VGG models are obtained, and we realized the importance of quantization for image inference.
Appears in Collections:	[Graduate Institute of Computer Science and Information Engineering] Electronic Thesis & Dissertation

Files in This Item:

File	Description	Size	Format
index.html		0Kb	HTML	59	View/Open

社群 sharing

Loading...