Object Detection And Segmentation Using Optimized Deep Learning Algorithms

Abstract

The effect of deep learning on video analysis including action detection and recognition
was not that important due to the nature of the video data format and lack of annotations.
Additionally, training profound neural networks on large-scale video datasets is extremely
expensive in terms of computation power and cost. Low-level features such as intensity, orientation,
and motion will render object identification for each pixel in image can solve the problem to a
greater extent. The present article has reviewed few existing techniques adopted by earlier research
that are implemented to process video data with prominence to detect various object classes. First,
we proposed a semantic segmentation using deep learning technique that suggest exploring an
alternate technique for object detection. Assuming pixel-wise annotations, an encoder-decoder
network structure widely used for semantic video frame segmentation is built using 3D CNN to
predict the pixel-wise labels for each video frame. Second, the paper proposed two optimized
YOLO v2 and YOLO v3 object detection and classification algorithm to improve the efficiency of
object classification. Finally, the performance of the proposed algorithms is evaluated using an
experimental approach and shown that the proposed techniques proved their efficiency with 99% of
accuracy levels in object detection and segmentation when compared with the other classification
versions with 61%.