Search

Curing modulation is one of the important processes in tobacco production, so it is crucial to recognize tobacco flue-curing states effectively and accurately. This study created a dataset of the complete tobacco flue-curing process in a bulk curing barn environment and proposed a lightweight recognition model based on a feature skip connections module. Firstly, the image data was enhanced using a color correction matrix, which was used to recover the true color of the tobacco leaf in order to reduce the misidentification of adjacent states. Secondly, the convolutional neural network model proposed in this paper introduced Spatially Separable convolution to enhance the extraction of tobacco leaf texture features. Then, the standard convolution in Short-Term Dense Concatenate (STDC) was replaced with Depthwise Separable Convolutional blocks with different expansion rates to reduce the number of model parameters and FLOPs (Floating Point Operations Per Second). Finally, the Tobacco Flue-Curing State Recognition Network (TFSNet) was constructed by combining the SimAm attention mechanism. The experimental results showed that the model accuracy was improved by 1.63 percentage points after the color correction process. The recognition accuracy of TFSNet for the seven states of tobacco flue-curing was as high as 98.71%. The number of params and the FLOPs of the TFSNet model were 203,058 and 172.39 M, which were 98.18% and 90.55% lower than that of the ResNet18 model, respectively. The size of the model was 0.78 mb, and the time consumed per frame was only 21 ms. Compared with the mainstream model, TFSNet significantly improved the detection speed while maintaining high accuracy, and it provided effective technical support for the intelligentization of the tobacco flue-curing process.