A CMOS imaging device for visual prosthetics using on-pixel gray-scale erosion for edge detection.

摘要

This thesis proposes a new design for on-pixel implementation of gray-scale morphology, aimed at visual prosthetics. These implants rely on providing the patient with a sketch or a realistic edge representation of the real world. Edge detection can be carried out using classical algorithms or mathematical morphology.; Threshold decomposition of gray-scale images allows for the simplification of grayscale mathematical morphology into binary morphology. Binary morphology can be implemented on pixel by using AND, OR and inverter gates. Gray-scale morphology on the other hand, needs more complex operations. By using threshold decomposition, these operations can be carried out in terms of binary morphology. To carry out threshold decomposition in hardware, a specialized system is needed to decompose the gray-scale image into M binary images, where M is the largest value a pixel can have in the gray-scale image.; This thesis proposes a new method for gray-scale mathematical morphology, called bitwise decomposition of gray-scale images. The proposed method reduces the number of binary images produced from M to N, where N is the number of bits representing the pixels of a gray-scale image. Binary erosion can then be carried out on the generated binary images, and the eroded gray-scale image reconstructed.; Erosion was chosen for edge detection since it produces correctly placed edges in the edge image. Once the gray-scale eroded image is available, it is used in morphological edge detection, where the results are comparable with classical algorithms. Once testing the new method in software is complete, the method is implemented in hardware.; A pixel employing the method of bitwise decomposition is designed in 0.13 mum technology. The pixel is then placed in three array structures; a 4x4, 8x8 and a 16x16 array. All arrays produced correct results as expected.; Layout of the pixel is drawn and tested using the layout-level simulations. The results snatch that of schematic-level tests. The fill factor of the pixel came to 22%, and the device has a dynamic range of 46 dB. Pixel dimension is 17x17 mu m, and transistor count 107. An analog to digital converters resolution is 8-bits and has a conversion time of 40 musec. The power dissipation of the 16x16 is 0.38 mW.