Implementation of Real-Time Content-Based Image Retrieval with Hardware/Software Co-Design

[[abstract]]本論文利用軟/硬體共設計及驗證方式在可程式規劃邏輯元件上實現內容導向影像檢索。本系統使用廣義霍福轉換整合影像區塊表示影像內容，這個系統包含大量的區塊比對工作，而區塊比對流程所花費的時間比例佔整個內容導向影像檢索系統大部分，而且每個待比對的視覺物件的組成區塊(視覺樣式)均可獨立在資料庫影像中尋找其可能的相對區塊，所以它隱含著大量的細微平行化單元，很適合將此部分以硬體的方式實現。 在硬體架構方面，我們提出一使用二維陣列的處理單元(PE)平行結構設計，PE裡的每個計算單元(CU)則用以處理一個區塊比對工作，所有的CU同時進行比對動作，因此可大大地提高處理區塊比對的速度。在累加CU計算出的幾何參數票數部分，我們利用一區塊記憶體來儲存，PE處理完後，再從區塊記憶體裡找出最高票數的幾何參數，傳回給中央處理單元作影像相似度的計算。 本論文實現硬體時，主要是將區塊比對這部分的演算法，利用硬體描述語言來轉換成硬體結構，裡頭牽涉到的一些函數，是利用矽智財來處理，如此可加快設計速度。[[abstract]]In this thesis, a content based image retrieval system using hardware/software co-design on a programmable logic component is presented. In this work, we use generalize Hough transform (GHT) to integrate block information within an image for measuring similarity values. A large amount of block matching tasks are involved in the GHT. Actually, the block matching module spends most of the execution time for answering a query. In addition, each block (or visual pattern) of a visual query object can perform its block matching procedure to find the possible match blocks in database images individually. The system implicitly contains a large number of fine-grained parallel units. Thus, it is suitable to be implemented with a hardware architecture in order to achieve the goal of real-time image retrieval. In this work, we propose a systolic-like parallel architecture consisting of a two-dimensional array of process elements (PE). Each PE contains several computing units (CU) and each CU performs a block matching task. All the CUs do the block matching at the same time, so the retrieval speed of the system is fast. Furthermore, we use a block memory to accumulate votes of the global geometric transformation parameters in the GHT. The peaks of votes of the global geometric transformation parameters extracted from the block memory are finally transmitted to the master unit, which is used to calculate the similarity value between a query image and a database image and returns the relevant images to the user. In order to speed up the design and implement of the system, some silicon intellectual properties are used in our system