Abstract:

Objective The complexity of deep convolutional neural network models makes it difficult for embedded devices to meet real-time online detection, and this research works on an improved RFB-MobileNetV3(RFB-MNV3) method for cotton impurity detection.
Method Firstly, the MNV3 redundant network structure was reduced according to the construction of high-precision lightweight network model and the premise of ensuring high detection accuracy. Secondly, the 3×3 convolutional layer replaced the 5×5 convolutional layer and the 1×3+3×1 convolutional layer was folded to replace the 3×3 convolutional layer as the improved RFB module deployed to the pooling layer of the improved MNV3 to enhance the online detection speed and accuracy of cotton hash. Finally, the algorithm before and after the improvement and other detection algorithms were compared.
Results The influence of training times, different lighting changes and different camera shift poses on the model was investigated using the test set. The improved RFB-MNV3 network model was iteratively trained to improve the average accuracy of cotton impurity classification. The specific classification detection effect under the improved RFB-MNV3 model showed that the detection accuracy was 83%-96% as suggested by the average AP values of the detection results for each category, among which the best effect was achieved in identifying cotton seeds with 96% accuracy (Fig.11). The value of the improved RFB-MNV3 algorithm reached 88.15%, indicating that the accuracy and score (the score of impurity detection under the optimized algorithm) have reasonably high stability, i.e. the model can better classify cotton impurity detection and can basically meet the actual industrial production needs. The detection results were compared with those of the MNV3, YOLOv3, VGG16 and ResNet34 network models (Tab.2). The detection time of the improved RFB-MNV3 model reached 0.02 s, and the online detection accuracy of the improved RFB-MNV3 model reached 89.05%, which is 6.83% higher than MNV3 and 8.48%-17.32% higher than other network models. The average accuracy mean combined with the accuracy and recall rates can be utilized to evaluate the comprehensive performance of image classification. It can be seen that the improved RFB-MNV3 model has a mean accuracy value that is 6.31% higher than MNV3 and 8.76%-17.72% higher than other networks.
Conclusion The new network is improved on the basis of the MNV3 detection network, while the improved RFB-MNV3 module is combined to achieve the purpose of reducing the model parameters without basically losing the model accuracy, solving the problem that the complexity of the deep convolutional neural network model makes it difficult for the embedded device to meet the real-time online detection. The model proposed can effectively achieve the detection of lint images, while the model detection efficiency is high and the storage occupied is small, which can provide the necessary technical support for the development of embedded devices for lint image detection.

Key words: RFB-MobileNetV3, cotton impurity, online detection, network structure, lightweight model, image detection