Toward CNN Architectures for Image Detection

Currently, convolutional neural networks (CNNs) drive the state-of-the-art for object detection and classification in imagery. Pre-trained models exist with hundreds of million computational artificial neurons to classify images into hundreds of different classes. However, most image characterizatio...

Full description

Saved in:
Bibliographic Details
Main Authors: Israel, Steven A, Irvine, John M, Israel, Ieuan M
Format: Conference Proceeding
Language:English
Subjects:
classification
Computational modeling
convolutional neural networks
detection
Focusing
Image quality
information content
Keywords > machine learning
Neurons
Object detection
Throughput
Training
Online Access:Request full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Currently, convolutional neural networks (CNNs) drive the state-of-the-art for object detection and classification in imagery. Pre-trained models exist with hundreds of million computational artificial neurons to classify images into hundreds of different classes. However, most image characterization problems, detection and classification, require the discrimination across only a handful of classes. By focusing our training to the limited class list, we expect dramatically improved discrimination with lower resource costs. Ideally, these networks are trained using smaller datasets and needed far fewer computational neurons. As researchers focus more on the detection paradigm, the expected smaller, nimbler models that balance performance optimization between correctness and throughput. Our previous studies matched image quality with convolutional neural network parameters to develop strategies for automating orchestrated collection based upon information needs. Our current study extends those goals in two directions: 1. incorporate hidden layer depth; and 2. incorporate different number of output classes. In this study, we trained multiple CNNs and varied their window sizes, hidden layers, latent space dimensions, and output classes. The images were acquired from the rare planes database and have a variable size between 100 and 400 pixels on a side. The number of classes varied from 2 to 8. The images contain synthetic targets that are visually separable from the background. The outcome from these and previous experiments is to identify candidate parameters that enable more efficient image search models.
ISSN:2332-5615
DOI:10.1109/AIPR57179.2022.10092233