Design of an intelligent information assessment model based on multiscale convolutional neural networks within the realm of online education
Abstract
The utilization of a traditional Convolutional Neural Network, albeit with a single receptive field, leads to a diminished accuracy in image retrieval. Simultaneously, model training often suffers from overfitting when executed with a limited dataset. To address these challenges, this study introduces a multiscale-based Convolutional Neural Network for the creation of an intelligent information assessment model. A novel addition, the dilated Convolution, enhances the encoder by employing a four-layer dilated convolutional network to encode sample image features. It subsequently splices the deep features obtained from image extraction in both the training and test sets, facilitating multiscale feature group comparisons. Furthermore, the incorporation of a dense residual convolutional module, which incorporates dense connections, establishes multiple short-circuit connections between multilayer convolutional layers. This construct captures a wealth of contextual semantic information. To gauge the model's efficacy, experimental validation is conducted on two modest sample benchmark image datasets: Omniglot and Min-Imagenet. The results demonstrate an 86.6% accuracy rate, a 90.47% recall rate, an 88.71% precision rate, and an average reduction in model parameters and computational complexity of 12.3% and 33.12%, respectively. Additionally, the overall precision of the Overall Accuracy (OA) and the Kappa coefficient witnesses an average improvement of 81.95% and 0.804, respectively. These outcomes effectively affirm the algorithm's superiority in solving small-sample image classification problems. Following the analysis, it becomes evident that an effective intelligent information assessment model greatly facilitates the extraction and utilization of resources in the realm of online education, carrying significant practical implications for its development.