基于残差与注意力机制的原棉杂质检测

Detection of raw cotton impurities based on residual and attention mechanism

针对原棉杂质检测准确率不高的问题,以新疆棉花为研究对象,提出基于残差与注意力机制的原棉杂质检测算法。该算法为2阶段算法,准确率较高。首先,采集原棉杂质图象后对图像进行标注,再进行数据增广,可以避免训练过程中的过拟合现象,接着在原框架引入视觉注意力机制,通过改进算法结构来提高原棉杂质检测的准确率。其次,通过分析对比几种不同网络对原棉杂质检测的准确度,选取ResNet50为特征提取网络,该网络提高了算法的复杂特征提取能力。最后,采用RoI Align来减少量化误差,从而提高检测原棉杂质的准确性。实验结果表明,改进的算法虽然略微增多检测时间,但其整体检测准确率明显优于原算法,整体识别的准确率可达到94.84%,较改进前Faster R-CNN(faster region-based convolutional neural network)的识别率提高了5.58%。同时通过对比不同网络模型,结果显示改进后的Faster R-CNN对原棉杂质检测的效果更好。

Aiming at the problem of low accuracy of impurity detection in raw cotton, an improved algorithm based on residual and attention mechanism for detecting raw cotton impurity in Xinjiang cotton is proposed.The algorithm has high accuracy and is a two-stage algorithm.Firstly, the impurity images of raw cotton were collected and labeled, and then the data were enlarged to avoid the overfitting phenomenon in the training process.Then, visual attention mechanism is introduced into the original framework, and the accuracy of impurity detection of raw cotton is improved by advancing the algorithm structure.Secondly, by analyzing and comparing the accuracy of several different networks in detecting raw cotton impurities, ResNet50 was selected as the feature extraction network, which improved the complex feature extraction ability of the algorithm.Finally, ROI Align is used to reduce quantization errors and improve the accuracy of the detection of raw cotton impurities.Experimental results show that although the improved algorithm slightly increases the detection time, its overall detection accuracy is significantly better than the original algorithm, and the overall recognition accuracy can reach about 94.84%,which is 5.58% higher than the recognition rate of the faster region-based convolutional neural network(Faster R-CNN) before the improvement.Meanwhile, by comparing different network models, the results show that the improved Faster R-CNN has a better effect on the detection of raw cotton impurities.