the Identification Algorithm of Crayfish Body Features Based on the Improved Yolov8n Loss Function (Published)
The crayfish industry, primarily focused on Procambarus clarkii, is expanding rapidly but encounters challenges due to inadequate automation. Traditional manual visual inspection methods used for evaluating crayfish size and integrity during breeding and processing are labor-intensive and prone to errors. This study presents an improved algorithm based on the YOLOv8n framework to intelligently recognize and grade crayfish by accurately detecting the body, tail, and claws of Procambarus clarkii. The proposed approach introduces innovation by replacing the original loss function CIoU ( Complete Intersection over Union) with MPDIoU ( Modified Perfect Dark Intersection over Union). A novel scale factor, denoted as the ratio, has been introduced to adjust the size of the auxiliary bounding box within the loss calculation framework. This improvement, in conjunction with the MPDIoU loss function, notably enhances the accuracy and efficiency of bounding box regression. As a result, it enables the precise detection of the distinct body parts of crayfish, a pivotal advancement in automating the grading process. Empirical assessments demonstrated substantial enhancements in recognition accuracy.The incorporation of Inner-MPDIoU into the YOLOv8n model elevated the mean Average Precision ( mAP) from 83.7% to 90.8% across IoU thresholds ranging from 0.5 to 0.95. The results of this study highlight the effectiveness of the proposed algorithm in precisely recognizing critical elements of Procambarus clarkii. This investigation contributes to the overarching goal of attaining intelligent and accurate grading within the crayfish domain, potentially transforming conventional practices and enhancing industry productivity. The implications transcend mere automation, providing a groundwork for future exploration into intelligent systems tailored to the unique requirements of the crayfish industry.
Keywords: MPDIoU, Procambarus clarkia, YOLOv8, deep learning, image recognition
Advanced Simulation Datasets for Deep Learning-Based Photonic and Electromagnetic Research using FDTD Methods (Published)
We have provided finite numerical datasets using the FDTD technique, which describes the electromagnetic field distribution against the changes in material and structural characteristics in this paper. It holds information related to many numerical parameters and the field images of the corresponding shape and size for different configurations for Gold, MgF2, and glass. This dataset was created to enhance the study of photonics, optics, and electromagnetic waves and serve as an input for reinforcement learning models intended to make precise estimations of field behavior induced by material and/or geometrical inputs for photonics and optics. We also describe other datasets mentioned in the contextual literature and establish how our dataset is different by providing a more comprehensive, parameterized set of images and simulation data. Thus, describing the approach used to create the dataset, we discuss its possible use in various disciplines – from nanophotonic to machine learning where precise electromagnetic field modeling is needed.
Keywords: deep learning, electromagnetic field distribution, finite-difference time-domain (FDTD), parameterized dataset, photonics and plasmonics