Using Improved DeepLabV3+ for Complex Scene Segmentation

This paper presents a lightweight convolutional neural network model based on DeepLabV3+. By collecting and annotating an image dataset, sufficient sample support is provided for the model’s training. In order to enhance the model’s training speed and applicability, this study proposes replacing the backbone network Xception with MobileNetv2. This successfully reduces the number of model parameters, making the training process more rapid and efficient. Additionally, by refining the model structure and incorporating Coordinate Attention (CA) and Cascaded Feature Fusion (CFF) for feature integration, the feature map processing is further optimized. This allows the model to more accurately identify in complex background environments. Verification shows that the improved model performs excellently in metrics such as mIoU, mPA, mPrecision, and mRecall. During testing, mIoU reached 82.66%, mPA reached 88.20%, mPrecision reached 92.26%, and mRecall reached 87.67%. These results fully demonstrate the feasibility, accuracy, and robustness of the proposed method, providing strong support in practical applications.