import cv2 import numpy as np import math import torch import random from PIL import Image from torch.utils.data import DataLoader from torchvision.transforms import Resize torch.manual_seed(12345) random.seed(12345) np.random.seed(12345) device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu") class WireframeExtractor: def __call__(self, image: np.ndarray): """ Extract corners of wireframe from a barnacle image :param image: Numpy RGB image of shape (W, H, 3) :return [x1, y1, x2, y2] """ h, w = image.shape[:2] imghsv = cv2.cvtColor(image, cv2.COLOR_RGB2HSV) hsvblur = cv2.GaussianBlur(imghsv, (9, 9), 0) lower = np.array([70, 20, 20]) upper = np.array([130, 255, 255]) color_mask = cv2.inRange(hsvblur, lower, upper) invert = cv2.bitwise_not(color_mask) contours, _ = cv2.findContours(invert, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_NONE) max_contour = contours[0] largest_area = 0 for index, contour in enumerate(contours): area = cv2.contourArea(contour) if area > largest_area: if cv2.pointPolygonTest(contour, (w / 2, h / 2), False) == 1: largest_area = area max_contour = contour x, y, w, h = cv2.boundingRect(max_contour) # return [x, y, x + w, y + h] return x,y,w,h wireframe_extractor = WireframeExtractor() def show_anns(anns): if len(anns) == 0: return sorted_anns = sorted(anns, key=(lambda x: x['area']), reverse=True) ax = plt.gca() ax.set_autoscale_on(False) polygons = [] color = [] for ann in sorted_anns: m = ann['segmentation'] img = np.ones((m.shape[0], m.shape[1], 3)) color_mask = np.random.random((1, 3)).tolist()[0] for i in range(3): img[:,:,i] = color_mask[i] ax.imshow(np.dstack((img, m*0.35))) # def find_contours(img, color): # low = color - 10 # high = color + 10 # mask = cv2.inRange(img, low, high) # contours, hierarchy = cv2.findContours(mask, cv2.RETR_TREE, cv2.CHAIN_APPROX_SIMPLE) # print(f"Total Contours: {len(contours)}") # nonempty_contours = list() # for i in range(len(contours)): # if hierarchy[0,i,3] == -1 and cv2.contourArea(contours[i]) > cv2.arcLength(contours[i], True): # nonempty_contours += [contours[i]] # print(f"Nonempty Contours: {len(nonempty_contours)}") # contour_plot = img.copy() # contour_plot = cv2.drawContours(contour_plot, nonempty_contours, -1, (0,255,0), -1) # sorted_contours = sorted(nonempty_contours, key=cv2.contourArea, reverse= True) # bounding_rects = [cv2.boundingRect(cnt) for cnt in contours] # for (i,c) in enumerate(sorted_contours): # M= cv2.moments(c) # cx= int(M['m10']/M['m00']) # cy= int(M['m01']/M['m00']) # cv2.putText(contour_plot, text= str(i), org=(cx,cy), # fontFace= cv2.FONT_HERSHEY_SIMPLEX, fontScale=0.25, color=(255,255,255), # thickness=1, lineType=cv2.LINE_AA) # N = len(sorted_contours) # H, W, C = img.shape # boxes_array_xywh = [cv2.boundingRect(cnt) for cnt in sorted_contours] # boxes_array_corners = [[x, y, x+w, y+h] for x, y, w, h in boxes_array_xywh] # boxes = torch.tensor(boxes_array_corners) # labels = torch.ones(N) # masks = np.zeros([N, H, W]) # for idx in range(len(sorted_contours)): # cnt = sorted_contours[idx] # cv2.drawContours(masks[idx,:,:], [cnt], 0, (255), -1) # masks = masks / 255.0 # masks = torch.tensor(masks) # # for box in boxes: # # cv2.rectangle(contour_plot, (box[0].item(), box[1].item()), (box[2].item(), box[3].item()), (255,0,0), 2) # return contour_plot, (boxes, masks) # def get_dataset_x(blank_image, filter_size=50, filter_stride=2): # full_image_tensor = torch.tensor(blank_image).type(torch.FloatTensor).permute(2, 0, 1).unsqueeze(0) # num_windows_h = math.floor((full_image_tensor.shape[2] - filter_size) / filter_stride) + 1 # num_windows_w = math.floor((full_image_tensor.shape[3] - filter_size) / filter_stride) + 1 # windows = torch.nn.functional.unfold(full_image_tensor, (filter_size, filter_size), stride=filter_stride).reshape( # [1, 3, 50, 50, num_windows_h * num_windows_w]).permute([0, 4, 1, 2, 3]).squeeze() # dataset_images = [windows[idx] for idx in range(len(windows))] # dataset = list(dataset_images) # return dataset # def get_dataset(labeled_image, blank_image, color, filter_size=50, filter_stride=2, label_size=5): # contour_plot, (blue_boxes, blue_masks) = find_contours(labeled_image, color) # mask = torch.sum(blue_masks, 0) # label_dim = int((labeled_image.shape[0] - filter_size) / filter_stride + 1) # labels = torch.zeros(label_dim, label_dim) # mask_labels = torch.zeros(label_dim, label_dim, filter_size, filter_size) # for lx in range(label_dim): # for ly in range(label_dim): # mask_labels[lx, ly, :, :] = mask[ # lx * filter_stride: lx * filter_stride + filter_size, # ly * filter_stride: ly * filter_stride + filter_size # ] # print(labels.shape) # for box in blue_boxes: # x = int((box[0] + box[2]) / 2) # y = int((box[1] + box[3]) / 2) # window_x = int((x - int(filter_size / 2)) / filter_stride) # window_y = int((y - int(filter_size / 2)) / filter_stride) # clamp = lambda n, minn, maxn: max(min(maxn, n), minn) # labels[ # clamp(window_y - label_size, 0, labels.shape[0] - 1):clamp(window_y + label_size, 0, labels.shape[0] - 1), # clamp(window_x - label_size, 0, labels.shape[0] - 1):clamp(window_x + label_size, 0, labels.shape[0] - 1), # ] = 1 # positive_labels = labels.flatten() / labels.max() # negative_labels = 1 - positive_labels # pos_mask_labels = torch.flatten(mask_labels, end_dim=1) # neg_mask_labels = 1 - pos_mask_labels # mask_labels = torch.stack([pos_mask_labels, neg_mask_labels], dim=1) # dataset_labels = torch.tensor(list(zip(positive_labels, negative_labels))) # dataset = list(zip( # get_dataset_x(blank_image, filter_size=filter_size, filter_stride=filter_stride), # dataset_labels, # mask_labels # )) # return dataset, (labels, mask_labels) # from torchvision.models.resnet import resnet50 # from torchvision.models.resnet import ResNet50_Weights # print("Loading resnet...") # model = resnet50(weights=ResNet50_Weights.IMAGENET1K_V2) # hidden_state_size = model.fc.in_features # model.fc = torch.nn.Linear(in_features=hidden_state_size, out_features=2, bias=True) # model.to(device) # model.load_state_dict(torch.load("model_best_epoch_4_59.62.pth", map_location=torch.device(device))) # model.to(device) from segment_anything import sam_model_registry, SamAutomaticMaskGenerator, SamPredictor model = sam_model_registry["default"](checkpoint="./sam_vit_h_4b8939.pth") model.to(device) predictor = SamPredictor(model) mask_generator = SamAutomaticMaskGenerator(model) import gradio as gr import matplotlib.pyplot as plt import io def check_circularity(segmentation): img_u8 = segmentation.astype(np.uint8) im_gauss = cv2.GaussianBlur(img_u8, (5, 5), 0) ret, thresh = cv2.threshold(im_gauss, 0, 255, cv2.THRESH_BINARY) contours, hierarchy = cv2.findContours(thresh, cv2.RETR_TREE, cv2.CHAIN_APPROX_SIMPLE) con = contours[0] perimeter = cv2.arcLength(con, True) area = cv2.contourArea(con) if perimeter != 0: circularity = 4*math.pi*(area/(perimeter*perimeter)) if 0.8 < circularity < 1.2: return True else: return circularity def count_barnacles(image_raw, split_num, progress=gr.Progress()): progress(0, desc="Finding bounding wire") # crop image # h, w = raw_input_img.shape[:2] # imghsv = cv2.cvtColor(raw_input_img, cv2.COLOR_RGB2HSV) # hsvblur = cv2.GaussianBlur(imghsv, (9, 9), 0) # lower = np.array([70, 20, 20]) # upper = np.array([130, 255, 255]) # color_mask = cv2.inRange(hsvblur, lower, upper) # invert = cv2.bitwise_not(color_mask) # contours, _ = cv2.findContours(invert, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_NONE) # max_contour = contours[0] # largest_area = 0 # for index, contour in enumerate(contours): # area = cv2.contourArea(contour) # if area > largest_area: # if cv2.pointPolygonTest(contour, (w / 2, h / 2), False) == 1: # largest_area = area # max_contour = contour # x, y, w, h = cv2.boundingRect(max_contour) # image = cv2.cvtColor(image_raw, cv2.COLOR_BGR2RGB) # image = Image.fromarray(image_raw) # image = image[:,:,::-1] # image = image_raw # print(image.shape) # print(type(image)) # print(image.dtype) # print(image) corners = wireframe_extractor(image_raw) print(corners) # (0, 0, 1254, 1152) cropped_image = image_raw[corners[1]:corners[3]+corners[1], corners[0]:corners[2]+corners[0], :] print(cropped_image.shape) # cropped_image = cropped_image[100:400, 100:400] # print(cropped_image) # progress(0, desc="Generating Masks by point in window") # # get center point of windows # predictor.set_image(image) # mask_counter = 0 # masks = [] # for x in range(1,20, 2): # for y in range(1,20, 2): # point = np.array([[x*25, y*25]]) # input_label = np.array([1]) # mask, score, logit = predictor.predict( # point_coords=point, # point_labels=input_label, # multimask_output=False, # ) # if score[0] > 0.8: # mask_counter += 1 # masks.append(mask) # return mask_counter split_num = 2 x_inc = int(cropped_image.shape[0]/split_num) y_inc = int(cropped_image.shape[1]/split_num) startx = -x_inc mask_counter = 0 good_masks = [] centers = [] for r in range(0, split_num): startx += x_inc starty = -y_inc for c in range(0, split_num): starty += y_inc small_image = cropped_image[starty:starty+y_inc, startx:startx+x_inc, :] # plt.figure() # plt.imshow(small_image) # plt.axis('on') masks = mask_generator.generate(small_image) for mask in masks: circular = check_circularity(mask['segmentation']) if circular and mask['area']>500 and mask['area'] < 10000: mask_counter += 1 # if cropped_image.shape != image_raw.shape: # add_to_row = [False] * corners[0] # temp = [False]*(corners[2]+corners[0]) # temp = [temp]*corners[1] # new_seg = np.array(temp) # for row in mask['segmentation']: # row = np.append(add_to_row, row) # new_seg = np.vstack([new_seg, row]) # mask['segmentation'] = new_seg good_masks.append(mask) box = mask['bbox'] centers.append((box[0] + box[2]/2 + corners[0] + startx, box[1] + box[3]/2 + corners[1] + starty)) progress(0, desc="Generating Plot") # Create a figure with a size of 10 inches by 10 inches fig = plt.figure(figsize=(10, 10)) # Display the image using the imshow() function # plt.imshow(cropped_image) plt.imshow(image_raw) # Call the custom function show_anns() to plot annotations on top of the image # show_anns(good_masks) for coord in centers: plt.scatter(coord[0], coord[1], marker="x", color="red", s=32) # Turn off the axis plt.axis('off') # Get the plot as a numpy array # buf = io.BytesIO() # plt.savefig(buf, format='png', bbox_inches='tight', pad_inches=0) # buf.seek(0) # img_arr = np.frombuffer(buf.getvalue(), dtype=np.uint8) # buf.close() # # Decode the numpy array to an image # annotated = cv2.imdecode(img_arr, 1) # annotated = cv2.cvtColor(annotated, cv2.COLOR_BGR2RGB) # # Close the figure # plt.close(fig) # return annotated, mask_counter, centers return fig, mask_counter, centers # return len(masks) # progress(0, desc="Resizing Image") # cropped_img = raw_input_img[x:x+w, y:y+h] # cropped_image_tensor = torch.transpose(torch.tensor(cropped_img).to(device), 0, 2) # resize = Resize((1500, 1500)) # input_img = cropped_image_tensor # blank_img_copy = torch.transpose(input_img, 0, 2).to("cpu").detach().numpy().copy() # progress(0, desc="Generating Windows") # test_dataset = get_dataset_x(input_img) # test_dataloader = DataLoader(test_dataset, batch_size=1024, shuffle=False) # model.eval() # predicted_labels_list = [] # for data in progress.tqdm(test_dataloader): # with torch.no_grad(): # data = data.to(device) # predicted_labels_list += [model(data)] # predicted_labels = torch.cat(predicted_labels_list) # x = int(math.sqrt(predicted_labels.shape[0])) # predicted_labels = predicted_labels.reshape([x, x, 2]).detach() # label_img = predicted_labels[:, :, :1].cpu().numpy() # label_img -= label_img.min() # label_img /= label_img.max() # label_img = (label_img * 255).astype(np.uint8) # mask = np.array(label_img > 180, np.uint8) # contours, hierarchy = cv2.findContours(mask, cv2.RETR_TREE, cv2.CHAIN_APPROX_SIMPLE)\ # gt_contours = find_contours(labeled_input_img[x:x+w, y:y+h], cropped_img, np.array([59, 76, 160])) # def extract_contour_center(cnt): # M = cv2.moments(cnt) # cx = int(M['m10'] / M['m00']) # cy = int(M['m01'] / M['m00']) # return cx, cy # filter_width = 50 # filter_stride = 2 # def rev_window_transform(point): # wx, wy = point # x = int(filter_width / 2) + wx * filter_stride # y = int(filter_width / 2) + wy * filter_stride # return x, y # nonempty_contours = filter(lambda cnt: cv2.contourArea(cnt) != 0, contours) # windows = map(extract_contour_center, nonempty_contours) # points = list(map(rev_window_transform, windows)) # for x, y in points: # blank_img_copy = cv2.circle(blank_img_copy, (x, y), radius=4, color=(255, 0, 0), thickness=-1) # print(f"pointlist: {len(points)}") # return blank_img_copy, len(points) demo = gr.Interface(count_barnacles, inputs=[ gr.Image(type="numpy", label="Input Image"), ], outputs=[ # gr.Image(type="numpy", label="Annotated Image"), gr.Plot(label="Annotated Image"), gr.Number(label="Predicted Number of Barnacles"), gr.Dataframe(type="array", headers=["x", "y"], label="Mask centers") # gr.Number(label="Actual Number of Barnacles"), # gr.Number(label="Custom Metric") ]) # examples="examples") demo.queue(concurrency_count=10).launch()