Fish Image Segmentation Algorithm (FISA) for Improving the Performance of Image Retrieval System

Abstract

Abstract-The image features (local, global) pay vital role in image retrieval system. The effectiveness of these image features depends on the application domain, i.e., in some domains the global features generate better results while in others the local features give good results. Different species of fishes have different color, texture, and shape features in their body parts (head, abdomen, and tail). Previously most of the work, in fish image domain has been done using global features. This work claims that fish image retrieval system using local features can generate better results as compared to global features. This is because of the fact that fish image has different features in its body parts. In this research, a fish image segmentation algorithm is proposed to extract fish object from its background and then separate fish object into three distinguished body parts, i.e. head, abdomen, and tail. The proposed algorithm was tested on a subset of "QUT_fish_data" data set containing 369 fishes of various sizes of 30 species. The experimental results showed an accuracy of 87.5% on fish image segmentation and demonstrated the effectiveness of local features over global features.

Robust features extraction for general fish classification

“…Fish recognition is very complicated and difficult task but is useful to business and agriculture. Distortion, overlap, noise, distortion, occlusion, and also error in segmentation are among the challenges faced in achieving accurate and reliable fish recognition [21][22][23][24][25][26][27][28][29][30][31][32][33][34]. Among the works relevant to this study is one from Mokti and Salam [35].…”

Sashimi: A toolkit for facilitating high‐throughput organismal image segmentation using deep learning

“…Our software implements Matterport's (Abdulla 2017) Mask R-CNN ) implementation with a custom fish segmentation model trained using transfer learning (Razavian et al 2014) against the 328,000 image COCO dataset (Lin et al 2014). We focus on fish because they present a wide gamut of phenotypes and color and pattern diversity (Marshall 2000;Losey et al 2003;Marshall et al 2003a;Marshall et al 2003b;Salis et al 2018;Alfaro et al 2019;Salis et al 2019), and because machine learning approaches have recently been applied to the problem of identifying and measuring fishes (e.g., Yao et al 2013;Qin et al 2016;Baloch et al 2017;Garcia et al 2020;Yu et al 2020). Our toolkit provides five key contributions: (1) "plug-and-play" reproducible, automated segmentation of fish images in complex backgrounds directly out-of-the-box, (2) a central database for sharing and distributing custom-trained segmentation models of other underrepresented organisms to use within the toolkit, (3) the ability to quickly specify an organism of interest to segment from the animal classes already included in COCO without needing to modify the codebase, (4) additional image preprocessing tools for popular color pattern analysis workflows, such as colordistance (Weller & Westneat 2019), pavo (Maia et al 2013;Maia et al 2019), or patternize (Van Belleghem et al 2018), and (5) built-in qualitative and quantitative image segmentation accuracy and diagnostic tools directly compatible with the segmentation outputs from Sashimi.…”