Browsing by Issue Date, starting with "2020-06-08"
Now showing 1 - 1 of 1
Results Per Page
Sort Options
- Realtime image noise reduction FPGA implementation with edge detectionPublication . Jardim, Ricardo Jorge Ferreira; Dias, Fernando Manuel Rosmaninho Morgado FerrãoThe purpose of this dissertation was to develop and implement, in a Field Programmable Gate Array (FPGA), a noise reduction algorithm for real-time sensor acquired images. A Moving Average filter was chosen due to its fulfillment of a low demanding computational expenditure nature, speed, good precision and low to medium hardware resources utilization. The technique is simple to implement, however, if all pixels are indiscriminately filtered, the result will be a blurry image which is undesirable. Since human eye is more sensitive to contrasts, a technique was introduced to preserve sharp contour transitions which, in the author’s opinion, is the dissertation contribution. Synthetic and real images were tested. Synthetic, composed both with sharp and soft tone transitions, were generated with a developed algorithm, while real images were captured with an 8-kbit (8192 shades) high resolution sensor scaled up to 10 × 103 shades. A least-squares polynomial data smoothing filter, Savitzky-Golay, was used as comparison. It can be adjusted using 3 degrees of freedom ─ the window frame length which varies the filtering relation size between pixels’ neighborhood, the derivative order, which varies the curviness and the polynomial coefficients which change the adaptability of the curve. Moving Average filter only permits one degree of freedom, the window frame length. Tests revealed promising results with 2𝑛𝑛𝑛𝑛 and 4𝑡𝑡ℎ polynomial orders. Higher qualitative results were achieved with Savitzky-Golay’s better signal characteristics preservation, especially at high frequencies. FPGA algorithms were implemented in 64-bit integer registers serving two purposes: increase precision, hence, reducing the error comparatively as if it were done in floating-point registers; accommodate the registers’ growing cumulative multiplications. Results were then compared with MATLAB’s double precision 64-bit floating-point computations to verify the error difference between both. Used comparison parameters were Mean Squared Error, Signalto-Noise Ratio and Similarity coefficient.