An optimized domain-specific shrimp detection architecture integrating conditional GAN and weighted ensemble learning

Deep learning primarily operates on images which contain hidden patterns that are quantified through pixel intensities. Deep learning is used to analyze the image patterns and to recognize the objects. The detection process includes the creation of labels with bounding boxes, and it will be evaluated by using accuracy scores. Sometimes, there is a need to improve the accuracy score by changing the fine-tuning parameters or generating the synthetic data, which leads to reducing the gap in organizing the patterns. To address this, our research introduces the synthetic data generation for "enhanced shrimp detection using integrated augmentation (ESDIA)" approach to detect shrimps. The methodology is used to combine the shrimp images with different backgrounds in different phases: segmentation, dataset construction, and creating classifiers like faster recurrent convolution neural network (FRCNN) and you only look once (YOLOv7). The Enhanced Shrimp Detection algorithm generates the unique features set and also various parameters fetched from foundational classifiers. The segmentation phase can be done through grayscale conversion, edge detection, thresholding, morphological operations, and image compilation from myriad angles. To bolster our dataset volume and variance, the proposed system will generate the synthetic data with different variants of the backgrounds using generative adversarial networks. The precision of object (shrimp) detection rate can be gauged by the proposed model is witnessed a vital flow, with the mean average precision ranging from 80.53 to 89.13 which indicates the efficacy of ESDIA in elevating shrimp detection capabilities through DL paradigms.