[Objectives]Intelligent detection of diseased wheat kernels is important for the efficient, rapid, and accurate evaluation of wheat kernel quality. Existing deep neural network models for the classification of diseased wheat kernels have disadvantages such as large numbers of parameters and complexity of operations, which make it unsuitable for the deployment of the model on edge computing devices, thus affecting the efficiency of on-site classification of diseased wheat kernels. In this paper, a lightweight neural network algorithm for diseased wheat kernel classification was proposed. [Methods]In this study, the model was developed based on the lightweight network MobileNetV2 and added an improved CBAM(convolutional block attention module)attention mechanism. The improved model was fully integer quantized and deployed to mobile devices. Moreover, the proposed model was applied to classify four types of wheat kernels(fusarium-damaged kernels, common bunt of wheat kernels, broken kernels, and normal kernels). [Results]Compared to the previous MobileNetV2 network, the model combining the improved attention mechanism and MobileNetV2 network was improved, and the accuracy, precision, and recall rates for the model were improved by 3.15%, 3% and 3%, respectively. The improved model after full integer quantization achieved 99%, 94%, 99% and 96% recognition accuracy for fusarium-damaged kernels, common bunt of wheat kernels, broken kernels, and normal kernels, respectively. In addition, the size of the model was 2.36 MB, and the single inference time of this model at the edge computing device was 96.95 ms. [Conclusions]The improved algorithm of this paper has increased the model accuracy, reduced the size of the model, and accelerated the model inference speed. This study can provide guidance for the de-redundancy of fusarium-damaged kernel classification models. [ABSTRACT FROM AUTHOR]