Image detection for monitoring soft shell crab using internet of things

This study addresses the critical challenge of accurately detecting soft-shell crab in an open greenhouse aquaculture environment, where image distortion, glare, and reflections hinder reliable monitoring. The focus is on enhancing the quality of images for crab detection through advanced image rectification and enhancement techniques, integrated with the YOLOv7 object detection model. Crab detection is the first step of soft-shell crab detection. In soft-shell crab farming, the crab detection is labourintensive, requiring farmers to monitor crabs every two to three hours. By leveraging computer vision and deep learning technologies, this study aims to reduce manual labour and increase productivity. To tackle wide-angle lens distortions, Fisheye Image Correction and 2D-Image Flat-Field Correction are employed, reducing the uneven lighting and spatial distortion caused by the greenhouse setup. While these methods greatly enhance image clarity, significant challenges remain in mitigating reflections and glare that occur in the highly reflective greenhouse environment. Reflection reduction is achieved using the Non-Local Means (NLM) filter and 2D-Image Flat-Field Correction. The methodology involves capturing images using an IoT-based setup, where cameras mounted on rails monitor crabs in containers with 300 images taken. To address the limited dataset, data augmentation techniques are applied, which improve the training of the YOLOv7 model for more accurate detection. Experimental results show that data augmentation boosts detection performance by 7%, and the combination of Fisheyes Image Correction and Flat-Field Correction yields an overall detection precision of 90.3%, compared to 55.4% without image processing. This study demonstrates the effectiveness of the proposed methods, offering a robust solution for enhancing crab detection accuracy in greenhouse farming systems.