Life Jacket Wearing Recognition Based on Semantic Segmentation
Cheng Ye, Ji Li · 2022
In order to automatically identify whether a person wear life jacket, we propose an automatic detection model based on semantic segmentation. To gain a better result, we made improvements and optimizations on the basis of U-net, which is the most classic network in semantic segmentation. We replaced the max-pooling layer with a special convolutional layer, such a way can extract more features from feature map and can achieve the same effect of max-pooling. We added an extra dropout layer to prevent the overfitting in our model. We also used reflection padding in it. Reflection padding is used to make the convolution kernel to extract more features. We used binary cross entropy as loss function and Leaky Relu as activation function. Compared with the traditional semantic segmentation networks such as U-net and FCN, the precision of our improved model is about 2% higher than U-net, and the F1-measure is 2.6% higher than it. Our precision is 9% higher than FCN-32S and F1-measure is about 10% higher. The final accuracy of our model can achieve 93% and the IoU value is 4% higher than the traditional U-net. The last result can reach the level of human eyes, such an accuracy is sufficient to practical applications.