|
|
 |
| |
|
|
|
Cyclic Self-Organizing Map for Object Recognition |
|
|
|
PP: 1775-1788 |
|
|
|
doi:10.18576/isl/120523
|
|
|
|
Author(s) |
|
|
|
S. Anter,
H. K. Hussein,
M. Hashim Ali,
F. F. Mohamed,
|
|
|
|
Abstract |
|
|
| Object recognition is an important machine learning (ML) application. To have a robust ML application, we need three major steps: (1) preprocessing (i.e. preparing the data for the ML algorithms); (2) using appropriate segmentation and feature extraction algorithms to abstract the core features data and (3) applying feature classification or feature recognition algorithms. The quality of the ML algorithm depends on a good representation of the data. Data representation requires the extraction of features with an appropriate learning rate. Learning rate influences how the algorithm will learn about the data or how the data will be processed and treated. Generally, this parameter is found on a trial-and-error basis and scholars sometimes set it to be constant. This paper presents a new optimization technique for object recognition problems called Cyclic-SOM by accelerating the learning process of the self-organizing map (SOM) using a non-constant learning rate. SOM uses the Euclidean distance to measure the similarity between the inputs and the features maps. Our algorithm considers image correlation using mean absolute difference instead of traditional Euclidean distance. It uses cyclical learning rates to get high performance with a better recognition rate. Cyclic-SOM possesses the following merits: (1) it accelerates the learning process and eliminates the need to experimentally find the best values and schedule for the learning rates; (2) it offers one form of improvement in both results and training; (3) it requires no manual tuning of the learning rate and appears robust to noisy gradient information, different model architecture choices, various data modalities and selection of hyper-parameters and (4) it shows promising results compared to other methods on different datasets. Three wide benchmark databases illustrate the efficiency of the proposed technique: AHD Base for Arabic digits, MNIST for English digits, and CMU-PIE for faces.
|
|
|
|
|
|
|
|
