視網膜功能在積體電路上之實現與應用

Abstract

在這個論文裡，介紹了視網膜功能在積體電路上的實現與應用。論文的主體分為四個部分：(1) 對哺乳類動物視網膜完整的介紹；(2)回顧以雙載子接面電晶體為基礎之仿視網膜影像偵測器；(3) 回顧以雙載子接面電晶體為基礎之仿視網膜運動偵測器；(4) 新的神經型態晶片設計方法，以及模仿兔子視網膜內開暫態神經節細胞組的實驗晶片設計與結果。 哺乳類動物視網膜由五種不同的細胞構成，而每一種細胞又各自有不同的分類。除了將影像世界的訊息轉變為神經訊號外，視網膜還負責0.3%的大腦視覺處理功能。在本研究之初，我們嘗試以工程的角度去瞭解視網膜的功能，而這樣的瞭解也啟發了本論文的研究。 在以雙載子接面電晶體為基礎之仿視網膜影像偵測器的部分，首先提出了一個新型態的矽視網膜晶片，該矽視網膜晶片實現了脊椎動物視網膜的部分功能。在這個矽視網膜晶片裡，每個基礎單元都只包含兩個分開的感光雙載子接面電晶體。脊椎動物視網膜內水平細胞的平滑功能，由感光雙載子接面電晶體共用基極處，光線引發過量載子的分佈與再分佈動作有效實現。因此，這個矽視網膜晶片架構非常簡單而緊密，可以用金氧半場效電晶體場效電晶體製程技術將它做得很小。實驗結果也顯示這個矽視網膜晶片能夠萃取出影像的邊緣與偵測運動影像。接著介紹一個改良過的簡單而緊密的矽視網膜晶片架構。在這個架構中，被採用n通道金氧半場效電晶體場效電晶體連結不同基礎單元中的雙載子接面電晶體基極，來取代原來雙載子接面電晶體共用基極的平滑網路，形成獨特而緊密的架構，因此平滑網路的特性變得具有廣泛地可調性。此外，每個雙載子接面電晶體也包含一個額外的射極作為列開關，以此大幅降低基礎單元的面積並增加解析度。平滑網路的可調性與此晶片偵測動態影像的能力都經實驗證實，相信這個改良後的結構將非常適合用來實現超大型積體電路智慧型影像偵測器。在本章最後提出了一個低光電流金氧半場校電晶體視網膜焦平面感測器結構，該結構採用擬雙載子接面電晶體平滑網路與適應性施密特電流觸發器。這個結構非常簡單而緊密，而且可以工作在極低的電流範圍，而施密特電流觸發器的滯後作用可以根據產生的光電流適應性調整。量測結果顯示此視網膜焦平面感測器可以成功地應用在字元辨識掃描器的相關產品上。 在以雙載子接面電晶體為基礎之仿視網膜運動偵測器的部分，首先提出的是採用雙載子接面電晶體矽視網膜與跨越零點偵測器設計的二維速度與方向選擇性視覺感測晶片。在這個感測器當中，採用了以標記為基礎延遲與關連的演算法來偵測運動中影像的速度與方向。此外，此架構更採用二元的脈衝訊號作為關連訊號來增加速度與方向的選擇性。在此感測器中的每個偵測基本單元都有緊密的架構，其中包含了一個雙載子接面電晶體為基礎之仿視網膜單元、一個電流輸入的邊緣萃取器、兩個延遲路徑、跟四個關連器，經由實驗證明製作的感測器晶片運作無誤。接著提出一個可偵測任意運動的金氧半場效電晶體場效電晶體即時焦平面運動偵測器，此偵測器採用以關連器為基礎的新型演算法。在此設計當中，採用以雙載子接面電晶體為基礎之仿視網膜感光細胞與平滑網路來擷取影像，並提高影像的對比，同時採用以關連器為基礎的新型演算法來做訊號處理，以決定入射影像的速度與方向。運動速度與方向的計算誤差經由平均累積十六個取樣畫面後已大幅降低。實驗結果已成功證明此運動偵測器藉由調整影像擷取頻率，能偵測每秒一畫素到每秒140,000畫素的運動影像，而最小的可偵測位移為每採樣時間內移動5微米。也就是說，這個高性能新型運動偵測器能被應用在許多即時的運動偵測系統上。再下來提出了一個具有仿視網膜處理電路的角速度與方向選擇性的圓形運動偵測器的設計與分析，這個圓形運動偵測器採極性結構，並且對圓形運動的角速度與運動方向（順時針與逆時針）具有選擇性。這個感測器的設計依然採用以關連器為基礎的新型演算法，而每個相關的畫素都相隔45o 角。在關連畫素當中插入更多畫素有助於提高角速度的選擇性。角速度選擇性同時與影像邊緣的數目與位置有關，角速度與邊緣的關係在論文當中有詳細的分析 ，實驗的結果也與角速度與方向選擇性的分析結果互相印證。最後提出一個金氧半場效電晶體場效電晶體即時焦平面切變運動偵測器，採用以關連器為基礎的新型演算法來偵測局部運動向量，並採用以偽雙載子接面電晶體為基礎之仿視網膜處理電路來偵測並前處理影像。為了要偵測切變運動，偵測畫素安置在切變運動軌跡上，這個架構對選定的切變運動速度與方向具選擇性。切變運動選擇性以三種圖樣不同的運動速度加以驗證。 在本論文的最後部分，提出了一個新的設計方法用以設計金氧半場效電晶體場效電晶體神經型態晶片，並模擬了兔子視網膜的開暫態神經結細胞組。量測結果顯示，該 金氧半場效電晶體場效電晶體神經型態晶片的運作方式與生物量測結果相符，因此成功地驗證了該晶片的仿生功能。經由實驗驗證，這個設計方法所設計的神經型態晶片能協助解尚未揭露的視網膜細胞行為與視覺語言，而且也可用以設計所有的視網膜神經結細胞組。此外，該研究結果也使得許多極具潛力的視網膜晶片應用，諸如運動偵測、電腦視覺、人工視網膜跟生醫元件方面變得更有可行性。

In this thesis, implementations and applications of the retinal functions on integrated circuits are introduced. The main parts of this thesis include: (1) a complete introduction of the mammalian retina; (2) review of the BJT-based silicon retina image sensors; (3) review of the BJT-based silicon retina motion sensors; (4) a new design methodology used to implement CMOS neuromorphic chips and an experimental chip that imitates the ON brisk transient ganglion cell set of rabbits’ retinas. The mammalian retina is comprised of five different kinds of cells, for each kind can be divided into more types. Besides transducing the visual world into neural signals, these cells are in charge of 0.3% of the visual function of the brain. In the beginning of this research, we try to understand the retinal functions from the engineering point of view, thus inspiring the following research of the thesis. For the BJT-based silicon retina image sensors, firstly a new silicon retina is proposed to realize the functions of the vertebrate retina. In the proposed silicon retina, each basic cell consists of two separated bipolar phototransistors only. The smooth function of the horizontal cell in the vertebrate retina is efficiently achieved by the diffusion and redistribution of the photogenerated excess carriers in the common base region of the phototransistors. Thus, the structure of the new silicon retina is very simple and compact. It can be easily implemented in CMOS technologies with a small chip area. Experimental results show that the new silicon retina is capable of extracting the edge of the image and detecting the moving object. Then an improved BJT-based silicon retina with simple and compact structure is proposed and analyzed. In the proposed structure, the BJT smoothing network is implemented by placing enhancement n-channel MOSFET’s among the bases of parasitic BJT’s existing in a CMOS process to form a unique and compact structure. Thus, the smoothing characteristics can be tuned in a wide range. Moreover, an extra emitter is incorporated with each BJT at the pixel to act as the row switch. This reduces the cell area of the silicon retina and increases the resolution. The measurement results on the tunability of the smooth area in the smoothing network as well as the dynamic characteristics of the proposed silicon retina in detecting moving objects have been presented. It is believed that the improved structure is very suitable for the very large scale integration implementation of the retina and its application systems for CMOS smart sensors. Finally, a new structure of low-photocurrent CMOS retinal focal-plane sensor with pseudo-BJT smoothing network and adaptive current Schmitt trigger is proposed. The proposed structure is very simple and compact. Moreover, the proposed circuit could be operated for low-induced current levels (pA), and the current hysteresis of the proposed current Schmitt trigger could be adjusted adaptively according to the value of induced photocurrents. Measurement results show that the proposed new retinal focal-plane sensor has successfully been used in character recognition of scanner systems, such as pen scanners, etc. For the BJT-based silicon retina motion sensors, firstly a 2-D velocity- and direction-selective visual motion sensor with BJT-based silicon retina and temporal zero-crossing detector is proposed and implemented. In the proposed sensor, a token-based delay-and-correlate computational algorithm is adopted to detect the selected speed and direction of moving object images. Moreover, binary pulsed signals are used as correlative signals to increase the velocity and direction selectivities. Each basic detection cell in the sensor has a compact architecture which consists of one BJT-based silicon retina cell, one current-input edge extractor, two delay paths, and four correlators. The correct operations of the fabricated sensor chip have been verified through measurements. Then a CMOS real-time focal-plane motion sensor intended to detect the global motion, using BJT-based retinal smoothing network and the modified correlation-based algorithm, is proposed. In the proposed design, the BJT-based retinal photoreceptor and smoothing network are adopted to acquire image and enhance the contrast of an image while the modified correlation-based algorithm is used in signal processing to determine the velocity and direction of the incident image. The deviations of the calculated velocity and direction for different image patterns are greatly reduced by averaging the correlated output over 16 frame sampling periods. Experimental results have successfully confirmed that the proposed motion sensor can work with different incident images and detect a velocity between 1 pixel/sec and 140,000 pixels/sec via controlling the frame sampling period. The minimum detectable displacement in a frame sampling period is 5μm. Consequently, the proposed high-performance new motion sensor can be applied to many real-time motion detection systems. Later a CMOS angular velocity- and direction-selective rotation sensor with a retinal processing circuit is implemented and analyzed. The proposed rotation sensor has a polar structure and is selective of the angular velocity and direction (clockwise and counterclockwise) of the rotation of images. The correlation-based algorithm is adopted and each pixel in the rotation sensor is correlated with the pixel that is 45o apart. The angular velocity-selectivity is enhanced by placing more than one pixel between two correlated pixels. The angular velocity-selectivity is related to both the number and the positions of the edges in an image. Detailed analysis characterizes angular velocity-selectivity for different edges. The experimental results successfully verified the analyzed characteristics of angular velocity- and direction-selectivity. Finally, a CMOS focal-plane shear motion sensor is designed and implemented. The adopted motion computation method is based on the modified correlation-based algorithms to detect the local motion vectors. The adopted pseudo BJT-based retinal processing circuit is to sense and preprocess the incident image. In order to detect shear motion, the arrangement of pixels is along the path of shear motion in the chip of shear motion sensor. This structure is selective to the preferred shear direction and velocity of the images. The shear motion selectivity is verified after tested by three patterns translating at different velocities. In the last part of this thesis, a new design methodology is proposed to implement CMOS neuromorphic chips which imitate the ON brisk transient ganglion cell set of rabbits’ retinas. The measurement results on the fabricated CMOS neuromorphic chip are consistent with the biological measurement results. Thus the biological functions of the chip have been successfully verified. It can be used to understand more biological behaviors and visual language of retinas under different input optical images which have not yet been tested in biological experiments. Based on the results, the full ganglion cell sets of retina can be designed. Thus many potential applications of retinal chips on motion sensors, computer vision, retinal prosthesis, and biomedical devices are feasible.