二值化

threshold = 100 fig, axs = plt.subplots(1, 10, figsize=(20, 2)) for j in range(10): img = images[j].reshape(28, 28) binary = np.where(img > threshold, 1, 0) axs[j].imshow(binary, cmap='gray') axs[j].set_title(class_names[labels[j].item()]) plt.show()

边缘检测

edge_filter = np.array([[-1, -1, -1], [-1, 8, -1], [-1, -1, -1]]) fig, axs = plt.subplots(1, 10, figsize=(20, 2)) for j in range(10): img = images[j].reshape(28, 28) binary = np.where(img > threshold, 1, 0) edge = signal.convolve2d(binary, edge_filter, boundary='symm', mode='same') # 2维卷积运算 axs[j].imshow(edge, cmap='gray') axs[j].set_title(class_names[labels[j].item()]) plt.show()

降分辨率

downsample = nn.MaxPool2d(kernel_size=2, stride=2) img = images[0].reshape(1, 1, 28, 28) downsampled = downsample(img) fig, axs = plt.subplots(1, 2, figsize=(5, 2)) axs[0].imshow(img.reshape(28, 28), cmap='gray') axs[0].set_title('Original') axs[1].imshow(downsampled[0][0], cmap='gray') axs[1].set_title('Downsample') plt.show()

升分辨率

upsample = nn.Upsample(scale_factor=2, mode='nearest') img = images[0].reshape(1, 1, 28, 28) upsampled = upsample(img) fig, axs = plt.subplots(1, 2, figsize=(5, 2)) axs[0].imshow(img.reshape(28, 28), cmap='gray') axs[0].set_title('Original') axs[1].imshow(upsampled[0][0], cmap='gray') axs[1].set_title('Upsample') plt.show()

旋转与翻转

fig, axs = plt.subplots(1, 5, figsize=(10, 2)) img = images[0].reshape(28, 28) axs[0].imshow(img, cmap='gray') axs[0].set_title('Original') axs[1].imshow(np.rot90(img, k=1), cmap='gray') axs[1].set_title('Rotate 90') axs[2].imshow(np.rot90(img, k=2), cmap='gray') axs[2].set_title('Rotate 180') axs[3].imshow(np.rot90(img, k=3), cmap='gray') axs[3].set_title('Rotate 270') axs[4].imshow(np.flipud(img), cmap='gray') axs[4].set_title('Flip Up Down') plt.show()

fig, axs = plt.subplots(1, 5, figsize=(10, 2)) axs[0].imshow(img, cmap='gray') axs[0].set_title('Original') axs[1].imshow(np.fliplr(img), cmap='gray') axs[1].set_title('Flip Left Right') plt.show(