基于改进 EfficientNet 的轻量化小麦不完善粒识别模型

To address the problems of high complexity and difficult deployment of existing recognition models for imperfect wheat grains based on convolutional neural networks,a lightweight imperfect wheat grain recognition model ML − EfficientNet based on the improved EfficientNet − B0 was proposed. Firstly,a lightweight attention module LCSA was proposed by improving the CBAM attention module and replacing the SE module in the original network with the LCSA module,so that the model could obtain both channel information and spatial information to enhance the modelś recognition ability. Then,the structure of MBConv was adjusted by drawing on the CSPnet idea to realize the purpose of improving themodelś recognition accuracy while reducing the number of model parameters. Finally,the LCSA module was added after the first convolutional layer of the model to enhance the feature extraction capability of the model. The experimental results indicated that the recognition accuracy of the ML − EfficientNet model was 95. 71%,the number of parameters was 2. 863 M,and the floating point computation was 0. 376 G. Compared with the pre − improvement model,the recognition accuracy was improved by 1. 57 percentage points,the amount of parameters was reduced by 60%,and the amount of floating − point computation was reduced by 9%,effectively carrying out the recognition task of imperfect wheat grains and provided useful support for smart agriculture.