Image enhancement for detection of underwater moulted crabs in greenhouse soft-shell crab farming using deep learning

• Fisheye flattening surpasses stretch flattening in wide-angle distortion rectification • 2D-image flat field outperforms non-local mean filter for reflection removal • Data augmentation improves rectifications of wide-angle distortion and reflection • Counting algorithm is proposed to detect moulted crabs in soft-shell crab farming • Optimal methods successfully detect 92.0% of underwater moulted crabs in greenhouses In soft-shell crab farming, collecting the moulted crabs within a narrow window of one hour is crucial to meet the quality of high-grade paper shell crab. This paper incorporates IoT-based computer vision technology and deep learning techniques in greenhouse soft-shell crab farming to detect moulted crabs. The challenges arise in accurately detecting underwater crabs due to wide-angle distortion and reflections in images captured by cameras installed on rails within the greenhouse. Image enhancement methods are proposed to address these challenges and improve the performance of crab detection using the YOLOv7 deep-learning algorithm. Specifically, the performance of two methods for rectifying wide-angle distortion and two methods for removing reflections are compared, respectively. Data augmentation techniques are also applied to overcome the limitation of annotated underwater crab targets for deep-learning training. Results show that the fisheye flattening method outperforms the stretch flattening method for rectifying wide-angle distortion, while 2D-image flat field correction yields better results than the non-local mean filter for reducing reflections. Furthermore, data augmentation significantly enhances the performance of crab detection. The application of the best image enhancement methods and data augmentation techniques yields a successful detection rate of 92% for moulted crabs in greenhouse soft-shell crab farming.