An Image Dataset Quality Evaluation System for Industrial Object Detection Tasks

The performance of visual detection models in industrial applications is strongly influenced by training dataset quality. Although imaging scheme design and algorithm optimization are often emphasized, systematic dataset quality evaluation remains insufficient. To address this gap, this study proposes a dataset quality evaluation framework for industrial object detection. It includes four dimensions and thirteen quantifiable indices: three for acquisition environment, two for image quality, three for dataset scale, and five for annotation quality. Normalization based on theoretical maximum scores is used to reduce biases caused by different score ranges, and dimension weights are assigned using Taguchi orthogonal experiments. Validation is performed on five public and three self-constructed datasets using YOLOv12n and RT-DETR-R18. A positive correlation trend is observed between the proposed scores and detection accuracy, with PLCC/SRCC/Kendall's tau values of 0.685/0.850/0.764 and 0.656/0.826/0.691, respectively. After second-level weight optimization, the correlations increase to 0.775/0.922/0.837 and 0.748/0.898/0.764. Corresponding p-values and 95% confidence intervals are reported to quantify statistical uncertainty. Sensitivity analysis and ablation comparisons further verify the robustness and necessity of the proposed multidimensional framework. The proposed framework provides a quantifiable method and practical acquisition guidelines for improving industrial image dataset quality.