Machine Learning and AI Software Engineering
Niš, Central Serbia, Serbia
Content-based image retrieval utilizes representations of features that are automatically extracted from the
images, mainly perceptual properties, such as shape, color, texture, and spatial relationships to find similar-looking images.
The usage of CBIR techniques aids in resolving the problem of missing annotations from the extracted video frames that have more eminent noise and overall lower quality, by permitting them to inherit (to some extent and with lower confidence) the tags retrieved from the similar-looking frames, which will result in an overall increase of the trustworthiness of the object detection algorithms themselves.
This speech will review the implementation of a video content categorization model using the
YOLOv3 implementation pre-trained on the MS-COCO and fine-tuned on the custom dataset, followed by the presentation on the improvements in overall results by using the custom-built convolutional neural network for CBIR and tag suggestion.