PavanproJack / Fruit-Detection-in-Orchards

Fruit Detection and Counting Model research for Yield mapping and Robotic harvesting

Fruit Detection and Ripeness Analysis in Mango Orchards

Detection Results:

Yolov4 to the Left and Yolov5 to the Right

29th July 2020

YoloV4 and V5 models are trained and tested under consistent conditions of augmentations and pre-processing.

YoloV4 and YoloV5 Training Conditions:

Mangos are annotated in their respective formats which can be accessed from YoloV4 augmented Dataset and YoloV5 augmented Dataset folders.

Pre-Processing:

• Auto-orientation of pixel data
• Resizing to 416x416

Augmentation:

• 50% probability of horizontal flip
• 50% probability of vertical flip
• Random rotation of between -15 and +15 degrees
• Random shear of between -15° to +15° horizontally and -15° to +15° vertically
• Random brigthness adjustment of between 0 and +85 percent
• Random exposure adjustment of between -27 and +27 percent
• Salt and pepper noise was applied to 1 percent of pixels

Here are the results obtained with their best weights.

Parameters	Yolov4(6000 epochs)	Yolov5(300 epochs)	Yolov4-tiny(6000 epochs)	Detectron 2	Faster RCNN
Precision	0.76	0.5676	0.64
Recall	0.83	0.8427	0.68
F1-score	0.80	0.678	0.66
mAP@0.5	0.82	0.791	0.6146

References and Similar Work Results.

Title	Architectures Used	Data Augmentation Techniques	Evaluation of Detection Performance	Graphs Plotted	Results
Deep Fruit Detection in Orchards	Faster R-CNN	Image Flipping and Rescaling	Tiling approach FR-CNN	Average Precision Response Area under Precision Recall Curve, F1-score Avg.Precision vs Number of Training images	F1 score > 0.9; Precision = 0.958; Recall = 0.863
Fast implementation of real-time fruit detection in apple orchards usingdeep learning	LedNet	Two level scale amplification	Average Precision Response, IoU, Precision, Recall	Focal Loss and MSE functions vs Object confidence Score	Recall = 0.821; Accuracy = 0.853

Ripeness Analysis: