2024 · 19 citations · 10 references
ComputingComputer ScienceComputer VisionMachine LearningMachine VisionDeep LearningInstance SegmentationLoss FunctionObject DetectionObject RecognitionTransfer LearningBenchmark AnalysisDeep Learning-based ClassificationLoss ValidationPrecisionPrecision-recall TradeoffResilience
In computer vision applications such as object recognition and instance segmentation, deep learning techniques—more specifically, the Mask-RCNN architecture based on ResNet50—have shown exceptional capabilities. This work makes use of transfer learning to provide a thorough analysis of the ResNet50-Mask-RCNN model’s first implementation to understand its performance features and constraints on a range of datasets and situations. Additionally, the study presents benchmark data that will be used as a standard reference for the next comparative assessments that will concentrate on techniques with enhanced backbones. With significant convergence in validation loss values during training, validation results demonstrate the model’s efficacy in object identification and classification tasks. As shown by the Precision-Recall curve data, the model also achieves an impressive average mAP of 0.8095 and shows resilience with high recall rates at various precision levels. The AUC-ROC value of 0.9399 highlights how well the model can distinguish between positive and negative instances. High detection rates for the target class were noted in the testing results, which support the model’s effectiveness.
10
Transfer learning using computational intelligence: A survey
Jie Lü, Vahid Behbood, Hao Peng et al. · Knowledge-Based Systems · 2015 · 981 citations · Full text
Pre-trained models: Past, present and future
Xu Han, Zhengyan Zhang, Ning Ding et al. · AI Open · 2021 · 938 citations · Full text
Data augmentation: A comprehensive survey of modern approaches
Alhassan Mumuni, Fuseini Mumuni · Array · 2022 · 800 citations · Full text
Pasi Fränti, Radu Mariescu-Istodor · Pattern Recognition Letters · 2023 · 83 citations · Full text