Batch Target Recognition Count under Complex Conditions Based on Att-Unet

As we all know, the large number of counts is a challenging and time consuming task subject because of oversized number and complex conditions. However, the development of deep learning makes deep learning models very competitive in image segmentation. In this paper, we take cigarette filter rods as the research object. we first evaluate the standard Unet for the filter rod target recognition to separate target and background. Secondly, we use the focal loss function instead of the traditional cross-entropy function to solve the problem of imbalance between target and background area. Thirdly, we add a self-attention module in the traditional Unet convolutional layer to enhance the convolution effect. Fourth, we propose structural element detection criteria and round tangency matching strategy based on HMM (Hidden Markov Model) for the geometric relationship of filter rod position, which further improves the accuracy of the algorithm. We used Qu's [1], Mask-R-CNN [2], FCN [3], Deep-lab-V1 [4] and this paper’s algorithm to test the performance of 30000 images from the industrial site. The performance of this paper’s algorithm is completely better than the performance of the above algorithm.