Abstract: Aiming to address the limitations in current wheat unsound kernel recognition models that neglect the importance of extracting multi-scale features within the same convolutional layer, leading to low accuracy due to overly homogeneous features, a Multi-Scale Parallel Convolutional Neural Network (MSPCNeXt) was proposed. MSPCNeXt-v1 integrates dimensionality reduction and large kernel average decomposition, which not only reduces model overhead but also enhances representational capacity. MSPCNeXt-v2 further improves performance by splitting feature vectors along the channel dimension and feeding each part into separate branches. It utilizes factorization to convert large convolutional kernels into sequential orthogonal square kernels. Results showed the model’s effectiveness on the public dataset GrainSpace. Data indicated that compared with the baseline ConvNeXt, MSPCNeXt-v1 achieved an average accuracy increase of 2.605% and Top-1 accuracy increase of 2.353%, though it resulted in a larger model size. MSPCNeXt-v2 improves average accuracy by 2.297% and Top-1 accuracy by 1.912%, while also reducing computational load by 0.456 G and parameter count by 5.986 M. By extracting features of wheat unsound kernels through multi-scale parallel convolutions and leveraging convolutional decomposition and feature splitting to reduce model complexity, the model significantly enhances identification efficiency.