Enhancing EfficientNetV2 for Adult Age Estimation of Forensic Pelvis

Traditional age estimation is generally achieved by the grading and assignment of the bone sites, which is easily affected by the subjective factors of observers. To solve this problem, in this paper, deep learning technique is used, and a novel Enhancing EfficientNetV2 approach (EE) is proposed for adult age estimation of forensic pelvis. Firstly, a database is constructed including MSCT image data and 3D image data of 164 samples. Secondly, MSCT image data are preprocessed in a uniform size by resampled, and the key parts of 3D image data are intercepted and converted into a uniform size. Thirdly, an Enhancing EfficientNetV2 model is constructed with adding a linear layer and sigmoid active function in the end of EfficientNetV2 model for converting the feature extracted from the Enhancing EfficientNetv2 model into the predicted values and controlling the output values of this modeL in a specified range. Finally, our proposed EE approach can achieve better results with coefficient of determination (R<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> ) being 0.9336, and Root Mean Squared Error (RMSE) being 2.1231 by training on our constructed database. During the training, Mix-up is used for data augmentation and improving the generalization ability of the proposed model, and using Cosine learning rate decay is smoother in the later stages of training to facilitate more detailed optimization. Although deep learning technique is exploratory work with application of adult age estimation of forensic pelvis, numerical experiments are verified the effectiveness of the proposed enhancing EfficientNetV2 approach on our constructed database.