Erkan Okuyan

This project relates to a sub-class of deep learning applications called convolutional neural networks. It is specially designed to fasten processing speed for such applications, where processing is carried out over the central processing unit (CPU). Generally, fast co-processors are used for convolutional neural network applications. However, energy efficiency requirements for some embedded environments may necessitate the usage of CPUs.

An inherent sub-function is crucial in the fast execution of convolutional neural networks. This is a convolution operation, which can be thought of as passing a filter over an image. For this project, I have found an order of execution between pixels so that retrieved data to the cache is used more times before eviction from the cache. Thus, cache misses reduce, and overall calculation speed increases.

Acceleration achieved using this method greatly varies according to image sizes and cache structure or size of the underlying platform. We have achieved three to five times speedup for the limited embedded environments we have tested.

There is a pending patent application associated with this project (tr 2021/013858).

As this system has the capability of optical focus to different distances, the search algorithm for finding the best focusing distance needs to process images captured from various distances. Determining the order of distances to be processed is crucial so that execution finishes quickly. Furthermore, system-level constraints usually determine the distance for the next calculation step. These constraints typically arise because directional changes or sudden stops of the motors controlling the lens positions are costly. Thus, the order of distances to be processed determined by the search algorithm also has to respect system-level constraints.

I have formulated the focusing problem for this project with a dynamic programming approach as an effortless way to implement the longest increasing sub-sequence calculation. This formulation allows fast calculation without directional change and finds the optimal focus distance before much defocusing occurs after the optimal point, i.e., it provides an excellent early termination condition.

A granted patent application is associated with this project (PCT/TR2016/050155).

This project relates to a noise elimination method for detection applications utilizing a data structure capable of fast and efficient membership tests. Candidates obtained from each frame are tested with the data structure responsible for membership tests to see if the prior occurrence of the same candidate took place. If not, the candidate is processed. Otherwise, the candidate is marked as seen in the data structure. Since candidates produced by noise in the frame are more likely to be blocked by the membership tests than candidates belonging to the object of detection, the present method is expected to improve detection performance.

I have used a Bloom filter data structure to carry out fast membership tests. The Bloom filter supports O(1) time insertion and test operations as well as O(1) space requirements. This incredible feat is achieved by allowing a small percentage of false positives for memberships tests while never allowing false negatives or supporting deletions. Thus, I eliminated noisy data with minimal computational cost by using the Bloom filter structure.

A pending patent application is associated with this project (tr 2020/19736).