Fig. 9

The Local Detector consists of a convolutional layer a of size (32 × 3 × 3) followed by a batch normalization layer, a max pool of size 2 and ReLU activation. The same structure is repeated through c–d and e–f with convolutions of sizes (64 × 3 × 3) and (128 × 3 × 3) respectively. Then the output of (f) is rearranged as a vector of size 2048 in g) and followed by two fully connected layers of sizes 500 and 2