Combining DeepLabV3 + and LSTM for intelligent drying strategy optimization in fruits and vegetables based on appearance quality: A case study of Pleurotus eryngii

From the paper

Abstract

Based on the DeepLabV3 + semantic segmentation model, this research incorporated the Content-Aware Reassembly of Features (CARAFE) module, Concurrent Spatial and Channel Squeeze and Channel Excitation (scSE) attention mechanism, and Improved-MobileNetV2 backbone network, proposing the DeepLab-CSB model to detect the appearance quality of Pleurotus eryngii (P. eryngii) slices on-line during drying. The proposed model improved the IoU, MIoU, mAP, recall, and segmentation speed by 1.41, 0.74, 0.95, 1.90, and 16.97 percentage points respectively, compared to the DeepLabV3 + model. Additionally, based on the Long Short-Term Memory (LSTM), the drying rate, color difference value, and shrinkage rate ofP. eryngiislices were predicted in real time. Subsequently, the most favorable drying parameters for these three indicators were defined, realizing real-time control of the drying process. The LSTM’s R² value of all processes was greater than 0.9, while the RMSE and MAE values were both less than 0.1. By implementing this strategy for parameter control, drying experiments were conducted on P. eryngii slices, resulting in the highest comprehensive score (0.7274) for the dried slices. This innovative approach demonstrates a significant enhancement in the efficiency of hot-air drying for P. eryngii slices with high quality.