Skip to main content
Fig. 1 | Plant Methods

Fig. 1

From: Deep learning based high-throughput phenotyping of chalkiness in rice exposed to high night temperature

Fig. 1

Model Architecture. a A backbone CNN (e.g., ResNet-101) is trained to classify (resized) input grain images as chalky or non-chalky. ResNet-101 has four main groups of convolution layers, shown as Layer1, Layer2, Layer3, and Layer4, consisting of 3, 4, 23 and 3 bottleneck blocks, respectively. b Each bottleneck block starts and ends with a \(1\times 1\) convolution layer, and has a \(3\times 3\) layer in the middle. The number of filters in each layer is shown after the kernel dimension. c Grad-CAM uses the gradients of the chalky category to compute a weight for each feature map in a convolution layer. The weighted average of the features maps, transformed using the ReLU activation, is used as the heatmap for the current image at inference time

Back to article page