# All imports # Math import math import cv2 import numpy as np from scipy.stats import pearsonr # Torch import torch import torchvision.transforms as transforms from torchvision.datasets import Caltech101 from torchvision.models import resnet50, ResNet50_Weights # OS and env from os import getenv from dotenv import load_dotenv import warnings load_dotenv() # MongoDB from pymongo import MongoClient # Visualizing import matplotlib.pyplot as plt def getCollection(db, collection): """Load feature descriptor collection from MongoDB""" client = MongoClient("mongodb://localhost:27017") return client[db][collection] def datasetTransform(image): """Transform while loading dataset as scaled tensors of shape (channels, (img_shape))""" return transforms.Compose( [ transforms.ToTensor() # ToTensor by default scales to [0,1] range, the input range for ResNet ] )(image) def loadDataset(dataset): """Load TorchVision dataset with the defined transform""" return dataset( root=getenv("DATASET_PATH"), download=False, # True if you wish to download for first time transform=datasetTransform, ) dataset = loadDataset(Caltech101) class GridPartition: """Class transform to partition image into (rows, cols) grid""" def __init__(self, rows, cols): self.rows = rows self.cols = cols def __call__(self, img): # img is in (C,(H,W)) format, so first element is channel img_width, img_height = img.size()[1:] cell_width = img_width // self.cols cell_height = img_height // self.rows grids = [] for i in range(self.rows): for j in range(self.cols): left = j * cell_width right = left + cell_width top = i * cell_height bottom = top + cell_height # Slice out grid = img[:, left:right, top:bottom] grids.append(grid) return grids def compute_color_moments(grid_cell): """Compute color moments (mean, std. deviation, skewness), assuming RGB channels""" grid_cell = np.array(grid_cell) # Convert tensor to NumPy array moments = [] for channel in range(3): # Iterate over RGB channels channel_data = grid_cell[:, :, channel] mean = np.mean(channel_data) std_dev = np.std(channel_data) # Avoiding NaN values skew_cubed = np.mean((channel_data - mean) ** 3) if skew_cubed > 0: skew = math.pow(skew_cubed, float(1) / 3) elif skew_cubed < 0: skew = -math.pow(abs(skew_cubed), float(1) / 3) else: skew = 0 moments.append([mean, std_dev, skew]) return moments def compute_color_moments_for_grid(grid): color_moments = [compute_color_moments(grid_cell) for grid_cell in grid] return np.array(color_moments).flatten() def combine_color_moments(grid_color_moments): return torch.Tensor(grid_color_moments).view( 10, 10, 3, 3 ) # resize as needed: 10x10 grid, 3 channels per cell, 3 moments per channel # Transform pipeline to get CM10x10 900-dimensional feature descriptor CM_transform = transforms.Compose( [ transforms.Resize((100, 300)), # resize to H:W=100:300 GridPartition( rows=10, cols=10 ), # partition into grid of 10 rows, 10 columns as a list compute_color_moments_for_grid, combine_color_moments, ] ) def compute_gradient_histogram(grid_cell): """Compute HOG using [-1,0,1] masks for gradient""" histograms = [] # Convert grid cell to NumPy array grid_array = np.array(grid_cell, dtype=np.float32) grid_array = grid_array.reshape( grid_array.shape[1], grid_array.shape[2] ) # ignore extra dimension # Compute the gradient using first-order central differences dx = cv2.Sobel( grid_array, cv2.CV_32F, dx=1, dy=0, ksize=1 ) # first order x derivative = [-1, 0, 1] dy = cv2.Sobel( grid_array, cv2.CV_32F, dx=0, dy=1, ksize=1 ) # first order y derivative = [-1, 0, 1]^T # Compute magnitude and direction of gradients magnitude = np.sqrt(dx**2 + dy**2) direction = np.arctan2(dy, dx) * 180 / np.pi # in degrees # Compute HOG - 9 bins, counted across the range of -180 to 180 degrees, weighted by gradient magnitude histogram, _ = np.histogram(direction, bins=9, range=(-180, 180), weights=magnitude) histograms.append(histogram) return histograms def compute_histograms_for_grid(grid): histograms = [compute_gradient_histogram(grid_cell) for grid_cell in grid] return np.array(histograms).flatten() def combine_histograms(grid_histograms): return torch.Tensor(grid_histograms).view(10, 10, 9) # Transform pipeline to get HOG10x10 900-dimensional feature descriptor HOG_transform = transforms.Compose( [ transforms.Grayscale(num_output_channels=1), # grayscale transform transforms.Resize((100, 300)), # resize to H:W=100:300 GridPartition( rows=10, cols=10 ), # partition into grid of 10 rows, 10 columns as a list compute_histograms_for_grid, combine_histograms, ] ) def loadResnet(): """Load ResNet50 pre-trained model with default weights""" # Load model model = resnet50(weights=ResNet50_Weights.DEFAULT) # try to use Nvidia GPU if torch.cuda.is_available(): dev = torch.device("cuda") torch.cuda.empty_cache() else: dev = torch.device("cpu") model = model.to(dev) model.eval() # switch to inference mode - important! since we're using pre-trained model return model, dev model, dev = loadResnet() class FeatureExtractor(torch.nn.Module): """Feature extractor module for all layers at once""" def __init__(self, model, layers): super().__init__() self.model = model self.layers = layers self._features = {layer: None for layer in layers} # store layer outputs here # Create hooks for all specified layers at once for layer_id in layers: layer = dict(self.model.named_modules())[ layer_id ] # get actual layer in the model layer.register_forward_hook( self.save_outputs_hook(layer_id) ) # register feature extractor hook on layer # Hook to save output of layer def save_outputs_hook(self, layer_id): def fn(_module, _input, output): self._features[layer_id] = output return fn # Forward pass returns extracted features def forward(self, input): _ = self.model(input) return self._features def resnet_extractor(image): """Extract image features from avgpool, layer3 and fc layers of ResNet50""" resized_image = ( torch.Tensor(np.array(transforms.Resize((224, 224))(image)).flatten()) .view(1, 3, 224, 224) .to(dev) ) # Attach all hooks on model and extract features resnet_features = FeatureExtractor(model=model, layers=["avgpool", "layer3", "fc"]) features = resnet_features(resized_image) avgpool_2048 = features["avgpool"] # Reshape the vector into row pairs of elements and average across rows avgpool_1024_fd = torch.mean(avgpool_2048.view(-1, 2), axis=1) layer3_1024_14_14 = features["layer3"] # Reshape the vector into 1024 rows of 196 elements and average across rows layer3_1024_fd = torch.mean(layer3_1024_14_14.view(1024, -1), axis=1) fc_1000_fd = features["fc"].view(1000) return ( avgpool_1024_fd.detach().cpu().tolist(), layer3_1024_fd.detach().cpu().tolist(), fc_1000_fd.detach().cpu().tolist(), ) def get_all_fd(image_id, img=None, label=None): """Get all feature descriptors of a given image""" img_shape = np.array(img).shape if img_shape[0] >= 3: true_channels = 3 else: # stacking the grayscale channel on itself thrice to get RGB dimensions img = torch.tensor(np.stack((np.array(img[0, :, :]),) * 3, axis=0)) true_channels = 1 cm_fd = CM_transform(img).tolist() hog_fd = HOG_transform(img).tolist() avgpool_1024_fd, layer3_1024_fd, fc_1000_fd = resnet_extractor(img) return { "image_id": image_id, "true_label": label, "true_channels": true_channels, "cm_fd": cm_fd, "hog_fd": hog_fd, "avgpool_fd": avgpool_1024_fd, "layer3_fd": layer3_1024_fd, "fc_fd": fc_1000_fd, } def euclidean_distance_measure(img_1_fd, img_2_fd): img_1_fd_reshaped = img_1_fd.flatten() img_2_fd_reshaped = img_2_fd.flatten() # Calculate Euclidean distance return math.dist(img_1_fd_reshaped, img_2_fd_reshaped) def cosine_distance_measure(img_1_fd, img_2_fd): img_1_fd_reshaped = img_1_fd.flatten() img_2_fd_reshaped = img_2_fd.flatten() # Calculate dot product dot_product = np.dot(img_1_fd_reshaped, img_2_fd_reshaped.T) # Calculate magnitude (L2 norm) of the feature descriptor magnitude1 = np.linalg.norm(img_1_fd_reshaped) magnitude2 = np.linalg.norm(img_2_fd_reshaped) # Calculate cosine distance (similarity is higher => distance should be lower, so subtract from 1) cosine_similarity = dot_product / (magnitude1 * magnitude2) return 1 - cosine_similarity def pearson_distance_measure(img_1_fd, img_2_fd): # Replace nan with 0 (color moments) img_1_fd_reshaped = img_1_fd.flatten() img_2_fd_reshaped = img_2_fd.flatten() # Invert and scale in half to fit the actual range [-1, 1] into the new range [0, 1] # such that lower distance implies more similarity return 0.5 * (1 - pearsonr(img_1_fd_reshaped, img_2_fd_reshaped).statistic) valid_feature_models = ["cm", "hog", "avgpool", "layer3", "fc"] valid_distance_measures = { "euclidean": euclidean_distance_measure, "cosine": cosine_distance_measure, "pearson": pearson_distance_measure, } def show_similar_images( fd_collection, target_image_id, target_image=None, target_label=None, k=10, feature_model="fc", distance_measure=pearson_distance_measure, save_plots=False, ): """Set `target_image_id = -1` if giving image data and label manually""" assert ( feature_model in valid_feature_models ), "feature_model should be one of " + str(valid_feature_models) assert ( distance_measure in valid_distance_measures.values() ), "distance_measure should be one of " + str(list(valid_distance_measures.keys())) all_images = fd_collection.find() # if target from dataset if target_image_id != -1: print( "Showing {} similar images for image ID {}, using {} for {} feature descriptor...".format( k, target_image_id, distance_measure.__name__, feature_model ) ) # store distance to target_image itself min_dists = {target_image_id: 0} # in phase 2, we only have even-numbered image IDs in database if target_image_id % 2 == 0: # Get target image's feature descriptors from database target_image_fds = fd_collection.find_one({"image_id": target_image_id}) else: # Calculate target image's feature descriptors target_image, target_label = dataset[target_image_id] target_image_fds = get_all_fd(target_image_id, target_image, target_label) target_image_fd = np.array(target_image_fds[feature_model + "_fd"]) for cur_img in all_images: cur_img_id = cur_img["image_id"] # skip target itself if cur_img_id == target_image_id: continue cur_img_fd = np.array(cur_img[feature_model + "_fd"]) cur_dist = distance_measure( cur_img_fd, target_image_fd, ) # store first k images irrespective of distance (so that we store no more than k minimum distances) if len(min_dists) < k + 1: min_dists[cur_img_id] = cur_dist # if lower distance: elif cur_dist < max(min_dists.values()): # add to min_dists min_dists.update({cur_img_id: cur_dist}) # remove greatest distance by index min_dists.pop(max(min_dists, key=min_dists.get)) min_dists = dict(sorted(min_dists.items(), key=lambda item: item[1])) # Display the target image along with the k images fig, axs = plt.subplots(1, k + 1, figsize=(48, 12)) for idx, (img_id, distance) in enumerate(min_dists.items()): cur_img, _cur_label = dataset[img_id] axs[idx].imshow(transforms.ToPILImage()(cur_img)) if idx == 0: axs[idx].set_title(f"Target image") else: axs[idx].set_title(f"Distance: {round(distance, 3)}") axs[idx].axis("off") if save_plots: plt.savefig( f"Plots/Image_{target_image_id}_{feature_model}_{distance_measure.__name__}_k{k}.png" ) plt.show() # else, if target from some image file else: print( "Showing {} similar images for given image, using {} for {} feature descriptor...".format( k, distance_measure.__name__, feature_model ) ) # store distance to target_image itself min_dists = {-1: 0} target_image_fds = get_all_fd(-1, target_image, target_label) target_image_fd = np.array(target_image_fds[feature_model + "_fd"]) for cur_img in all_images: cur_img_id = cur_img["image_id"] cur_img_fd = np.array(cur_img[feature_model + "_fd"]) cur_dist = distance_measure( cur_img_fd, target_image_fd, ) # store first k images irrespective of distance (so that we store no more than k minimum distances) if len(min_dists) < k + 1: min_dists[cur_img_id] = cur_dist # if lower distance: elif cur_dist < max(min_dists.values()): # add to min_dists min_dists.update({cur_img_id: cur_dist}) # remove greatest distance by index min_dists.pop(max(min_dists, key=min_dists.get)) min_dists = dict(sorted(min_dists.items(), key=lambda item: item[1])) # Display the target image along with the k images fig, axs = plt.subplots(1, k + 1, figsize=(48, 12)) for idx, (img_id, distance) in enumerate(min_dists.items()): if idx == 0: axs[idx].imshow(transforms.ToPILImage()(target_image)) axs[idx].set_title(f"Target image") else: cur_img, _cur_label = dataset[img_id] axs[idx].imshow(transforms.ToPILImage()(cur_img)) axs[idx].set_title(f"Distance: {round(distance, 3)}") axs[idx].axis("off") if save_plots: plt.savefig( f"Plots/Image_{target_image_id}_{feature_model}_{distance_measure.__name__}_k{k}.png" ) plt.show()