A Practical Visual Perception-based Species Recognition System for Non-setae Marine Microalgae

Abstract

Marine microalgae play a vital role in maintaining the balance of global marine ecosystems and achieving carbon neutrality goals. In this paper, 21 species of common non-setae microalgae in China Sea are taken as the research object, and a photomicrograph classification system based on visual perception is designed to automatically identify the species of microalgae according to their biological morphological characteristics. The main work includes: (1) Extract the microalgal cell targets from the photomicrographs. Aiming at non-setae microalgae, a non-interactive GrabCut method combined with salient region detection is proposed to extract single or multiple cells. First, a boundary correlation-based salient region detection algorithm is employed to localize cell regions. Then the generated saliency map is pre-segmented, and the position information in the pre-segmentation result is used to construct a rectangular window and a mask image respectively, and use them to initialize the GrabCut algorithm. Then iteratively execute GrabCut. Remove noise and smooth cell edges by extracting larger contours and median filtering, respectively. During cell extraction, similarity detection is introduced to exclude pseudo-cell regions. (2) 19 shape features such as body ratio, eccentricity, and rectangularity of microalgal cells are extracted, and the Support Vector Machine is used as a classifier to perform pattern recognition on the feature data. The experimental results show that the recognition scheme proposed in this paper has achieved a relatively ideal recognition effect on the non-setae microalgae, with a recognition accuracy rate of 81.76%.