app.py

import cv2
import numpy as np
import math
import torch
import random

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")

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)

import gradio as gr


def count_barnacles(raw_input_img, labeled_input_img, 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)

    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(shape=(500, 500), type="numpy", label="Input Image"),
                        gr.Image(shape=(500, 500), type="numpy", label="Masked Input Image")
                    ],
                    outputs=[
                        gr.Image(shape=(500, 500), type="numpy", label="Annotated Image"),
                        gr.Number(label="Predicted Number of Barnacles"),
                        gr.Number(label="Actual Number of Barnacles"),
                        gr.Number(label="Custom Metric")
                    ])
                    # examples="examples")
demo.queue(concurrency_count=10).launch()