Sea Ice Images Classification Based on Optimized Deeplabv3+

This study proposes an optimized Deeplabv3+ network for sea ice image classification. Firstly, motivated by the advantage of MobileNetv2’s lightweight model structure, the Deeplabv3+ model is combined with the Mobilenetv2 backbone network, to accelerate the convergence speed of the model. Secondly, in the encoding stage, the efficient channel attention (ECA) mechanism is introduced, to strengthen the key channel features and weaken the irrelevant channel features, so that the final fused feature maps are not only semantically rich but also relatively fine. To test the effectiveness of the optimization model, the experimental results are evaluated by qualitative analysis and quantitative analysis. For the qualitative analysis, the optimized Deeplabv3+ model is validated on a test set with better sea ice image classification results, more complete sea ice contours, and smoother sea ice segmentation details. The quantitative analysis of the mean intersection over union (MIoU) evaluation metric shows that the optimized Deeplabv3+ is 4.15% higher than the traditional algorithm, representing a better performance of the optimized model.