Comparing the Performance of Pre-trained Deep Learning Models in Object Detection and Recognition

Abstract

The aim of this study is to evaluate the performance of the pre-trained models and compare them with the probability percentage of prediction in terms of execution time. This study uses the COCO dataset to evaluate both pre-trained image recognition and object detection, models. The results revealed that Tiny-YoloV3 is considered the best method for real-time applications as it takes less time. Whereas ResNet 50 is required for those applications which require a high probability percentage of prediction, such as medical image classification. In general, the rate of probability varies from 75% to 90% for the large objects in ResNet 50. Whereas in Tiny-YoloV3, the rate varies from 35% to 80% for large objects, besides it extracts more objects, so the rise of execution time is sensible. Whereas small size and high percentage probability makes SqueezeNet suitable for portable applications, while reusing features makes DenseNet suitable for applications for object identification.

Keywords

• deep learning
• image recognition
• object detection
• pre-trained models