MSPCNeXt: Multi-scale parallel convolution for identifying unsound kernels in stored wheat

Abstract

Recognizing unsound kernels in wheat is crucial for wheat storage. Most current techniques rely solely on single-scale convolutional neural networks, overlooking the importance of extracting multi-scale information in the same convolutional layer. This may lead to overly homogeneous extracted features and lower recognition accuracy. To solve this issue, we propose the Multi-scale Parallel Convolution Neural Network (MSPCNeXt). This model extracts image features by parallelizing convolutional kernels with different scales. The MSPCNeXt-Block comprises two versions, i.e., MSPCNeXt-v1 and MSPCNeXt-v2. After optimizing the parallel branching by using large kernel decomposition and feature segmentation, MSPCNeXt-v1 and MSPCNeXt-v2 ensure a broad receptive field while lowering the model complexity. We validate the effectiveness of MSPCNeXt on the public dataset GrainSpace. Compared with the baseline model ConvNeXt, MSPCNeXt-v1 improves the mean accuracy by 2.531% and Top-1 accuracy by 2.286%, but the model size is too large. MSPCNeXt-v2 improves the mean accuracy by 2.224% and Top-1 accuracy by 1.857%. Meanwhile, the computation amount is reduced by 0.456G and the number of parameters is reduced by 5.986M. In conclusion, MSPCNeXt provides a more efficient technical tool for sampling and testing of wheat storage and related processes.