Algorithm 1: Training process of ATT-MRCNN for the

Input: The collected orange picture data, selected COCO and VOC

Initialization.

Define the ATT-MRCNN.

Weights initialization using the COCO datasets with transfer learning manner.

Feature extraction.

Channel attention and Spatial attention mechanism for the feature map.

Model training: Epochs initialization.

while epoch<preset epoch do

Sample data from Input.

Feed data to the ATT-MRCNN.

Model updates.

End

Parameters Fine tuning

while epoch<preset epoch do

Validation dataset input.

Loss calculation.

Compute precision, recall and F1-score

Save the optimal model

end

Output: Trained ATT-MRCNN network