Fig. 2

A general overview of the approach we propose. Figure 2 shows the entire proposed system. Input images of any size are trained in our primary diagnostic unit, which generates boundary boxes and their location and category of infected areas in the image. The secondary diagnostic unit uses the bounding box as an input, and the secondary diagnostic unit independently trains the CNN filter bank to reduce the number of false positives generated by the primary unit. Both systems are further integrated into the level and location. K-means++ clusters the border shapes of training samples from PASCAL VOC and COCO datasets. a It shows that when k-means++ chooses different K values, k = 6 is chosen to balance the speed and the overlap rate of the IOU. b K-means + + clustering results show that thin and high boundaries account for the majority of the samples