Fast Object Detection in Light Field Imaging by Integrating Deep Learning with Defocusing

Abstract

Although four-dimensional (4D) light field imaging has many advantages over traditional two-dimensional (2D) imaging, its high computation cost often hinders the application of this technique in many fields, such as object detection and tracking. This paper presents a hybrid method to accelerate the object detection in light field imaging by integrating the deep learning with the depth estimation algorithm. The method takes full advantage of computation imaging of the light field to generate an all-in-focus image, a series of focal stacks, and multi-view images at the same time, and convolutional neural network and defocusing are consequently used to perform initial detection of the objects in three-dimensional (3D) space. The estimated depths of the detected objects are further optimized based on multi-baseline super-resolution stereo matching while efficiency is maintained, as well by compressing the searching space of the disparity. Experimental studies are conducted to demonstrate the effectiveness of the proposed method.

Self-supervision based super-resolution approach for light field refocused image

“…Digital Refocusing is an important application in LF, such as object detection and depth recovery 18,19,20,21 . We can reproject all the recorded light rays to a new plane with different image distances to form a new image plane, to achieve refocusing.…”

Binocular depth estimation method based on multi-layer guiding feature fusion

“…Stereo matching is a highly researched topic in the field of computer vision, with diverse applications such as object recognition and autonomous driving [1]. Traditional stereo matching methods encompass four key steps: matching cost computation, cost aggregation, disparity calculation, and disparity optimization [2].…”

Image Quality Evaluation of Light Field Image Based on Macro-Pixels and Focus Stack

“…Light field (LF) imaging technology is designed to record rich scenario information. Compared with ordinary two-dimensional (2D) images and binocular stereoscopic images, LF images are favored in researches like immersive stereoscopic display and object recognition because of their particular characteristics of dense view and post-focusing (Huang et al, 2016 ; Ren et al, 2017a ). For these applications, image quality degradation will directly affect the perception of the immersive experience and the accuracy of object recognition.…”