Object based segmentation and analysis using deep learning algorithm for cats and dogs images

• Main Author: N.Rajandhiran, Shreevishal
• Format: Final Year Project / Dissertation / Thesis
• Published: 2023
• Subjects: Q Science (General); T Technology (General)
• Online Access: http://eprints.utar.edu.my/5780/1/fyp_CS_2023_SNR.pdf; http://eprints.utar.edu.my/5780/

Image segmentation is the process of segmenting an image into the region of interest. These regions of interest (ROI) can be objects that are found in an image. Natural images consist of objects such as dogs, cats, and etc that can be segmented as region of interest. There are existing software such as Image Annotation Lab, ImageJ and Interactive Segmentation Tool that are used to delineate the ROI in natural images however, these software have common limitation. User intervention is constantly required to perform the segmentation. In addition, the algorithms used for the segmentation are conventional rather than modern deep learning techniques which is inevitably more efficient. Thus, this proposed work uses deep learning techniques to segment the region of interest in natural images, therefore reducing human intervention in performing the segmentation. The deep learning algorithm implemented is the Unet architecture and the Segnet architecture respectively. The dataset representing the narrative of natural images is the Oxford IIIT Pet Dataset which contains 7500 images of dogs and cats of different breeds with the respective ground truth images. The evaluation measurement to determine the performance of both implemented Unet architecture and Segnet architecture is the validation loss and the accuracy. The Unet model built from scratch has observed an accuracy of 89.22% , while the pretrained Unet model has observed an accuracy of 89.68% and the Segnet model built from scratch observed an accuracy of 84.87% in delineating the cats and dogs as ROIs in the natural images dataset.