DAM: Hierarchical Adaptive Feature Selection Using Convolution Encoder-Decoder Network for Strawberry Segmentation

Our Contributions

• We propose a single attention module (DAM) for channel and spatial attention and a parallel dilated convolution module (PDC) for aggregating multi-scale context.
• We validate the effectiveness of DAM and PDC by ample ablation experiments.
• We propose an optimal location for integrating our attention module in any existing network and compare results with other existing attention mechanisms.
• We propose a technique for visually interpreting segmentation networks by modifying Grad-CAM.
• A new dataset for the semantic segmentation of strawberries is introduced, consisting of four classes depending on the ripeness level of the fruit.

Paper Available Here