Contrast Mask RECALL Denoising LOGAND Color Mask EROSION Denoising Figure 14. this figure depicts the algorithmic steps of crosswalk detection on a CNN processor. in a way that the density of the black cells will be proportional to the darkness of the original image in each region. This complex operation can also be implemented by the execution of a single CNN template and it can also be found in the Template Library: R- 0.02 - 0.07 - 0.10 S S- 0.07 - 0.32 - 0.46 A = SS- 0.10 - 0.46 1.05 S- 0.07 - 0.32 - 0.46 S- 0.02 - 0.07 - 0.10 RT0.02 0.07 0.10 0.07 S S0.07 0.32 0.46 0.32 B = SS0.10 0.46 0.81 0.46 S0.07 0.32 0.46 0.32 S0.02 0.07 0.10 0.07 T Z=0 - 0.07 - 0.32 - 0.46 - 0.32 - 0.07 0.02VW 0.7 W 0.10WW 0.7 W 0.02W X - 0.02VW - 0 .7 W - 0.10WW - 0 .7 W - 0.02W X (6) This is a 5 # 5 template where the neighbourhood radius-introduced in Section II-equals 2. Another 3 # 3 variant of the halftoning template can also be found in the Template Library. C. Crosswalk Detection Not only simple operations but also complex and practical algorithms can be found in the Template Library. This collection contains not only operations but also series of operations and complex methods and algorithms. Here we will briefly describe an algorithm which can detect crosswalks in an urban environment. The algorithm was developed for visually impaired people and was tested on the large dataset of crosswalks [36]. 86 IEEE CIrCuItS aND SyStEmS magazINE Crosswalks consist of periodically changing regions of high/low contrast and high/low intensity. These changes are first detected by the algorithm on two parallel branches, by adaptive thresholding and color filtering operations. Later on the detections are filtered and objects which are not changing periodically and are either too small or too large are removed with a series of SMKILLER, HOLEFILLING and LOGDIF operations (which are templates from the Template Library). After this step the filtered high-contrast and high and low intensity regions are merged and the region is reconstructed using the original image. A depiction of the algorithm can be found on Fig. 14. A more detailed description of the algorithm along with all the operations can be found in the Template Library and in [36]. VIII. Research Trends As Moore's Law based device scaling along with performance scaling started to slow down, interest has continuously increased in new technologies and computational models to create new specialized architecture which enable the solution of given tasks faster and in a more energy efficient manner. CNNs are in this regard under investigation via the Semiconductor Research Corporation's benchmarking activities [37] as (i) they can solve a broad set of problems [5], and (ii) can leverage unique properties of both spin- and charge- based devices. The most commonly investigated problems are classification problems (along with regression and clusterization). In this description of future trends we will also limit the scope to classification. Various metrics SECOND quartEr 2018