Comparison of Region Proposal Methods for Marine Holograms

A novel technique to analyse deep-sea particles using digital holography (DH) and Raman spectroscopy (RS) is currently being developed. Since RS requires long measurement periods to provide reliable results in water, a flow control system is necessary to trap particles in the instrument. One way in which this can be achieved is to control the flow based on particles detected in DH images that can be rapidly acquired. In this paper, we explore the possibility of using a region proposal technique to detect possible object regions in holographic images. A model using a support vector machine (SVM) and the histogram of oriented gradients (HoG) has been developed. After being trained on a dataset consisting of 1589 particle and 947 non-particle regions, the model was tested on an image set of 10 independent images with 175 positive regions. It took around 0.5 second to analyse a 1000 × 760 image on average, and the proposed regions had a maximum overlap of 59.3% on average compared to the ground truth. The speed and accuracy of this method was also compared against 7 other possible methods. Amongst the tested methods, SVM was shown to provide the best solution for real-time region proposal on digital holograms. The results in this paper show that the SVM model using HoG features is a good solution to implement region proposal on digital holograms.