Development of Performance Optimized Rotation Tolerant Viola-Jones Based Blackbird Detection, a Throughput Optimized Asynchronous Mac Implementation, and Automated Wheat Lodging Estimation

Abstract

The research described in this doctoral dissertation focuses on three main topics:1) performance optimization of the Viola-Jones Algorithm (VJA) for red-winged blackbird (Agelaius phoeniceus) detection, 2) further increasing performance of an already optimized asynchronous Multiply and Accumulate (MAC) unit, and 3) development of a framework to differentiate between lodging and non-lodging areas of a field from visible and multispectral aerial drone images. The first topic explores VJA rotational robustness, since VJA object detection is inherently not invariant to in-plane object rotation. An efficient method to detect rotated blackbirds is developed, which provides a balance between detection accuracy and computational cost. The second topic further optimizes a previously developed high-speed asynchronous 72+32×32 MAC, which was the fastest in the literature, resulting in a speedup of 1.36 while also decreasing area by 8%. The third topic develops a model to distinguish lodging from non-lodging plots, using a Support Vector Machine model trained with color, texture, Normalized Difference Vegetation Index (NDVI), and height features. The model prediction accuracy is around 90%, indicating good performance in distinguishing lodging from non-lodging plots.