第八章 AI模型质量-3

import torchimport torch.nn.functional as Fimport torchvision.models as modelsimport torchvision.transforms as transformsimport cv2import numpy as npclass GradCAM:    def __init__(self, model, target_layer):        self.model = model        self.target_layer = target_layer        self.model.eval()        self.feature_maps = None        self.grads = None        self.hook_layers()    def hook_layers(self):        def forward_hook(module, input, output):            self.feature_maps = output.detach()        def backward_hook(module, grad_in, grad_out):            self.grads = grad_out[0].detach()        self.target_layer.register_forward_hook(forward_hook)        self.target_layer.register_backward_hook(backward_hook)    def __call__(self, input_tensor, target_class=None):        output = self.model(input_tensor)        if target_class is None:            target_class = torch.argmax(output)        self.model.zero_grad()        output[:, target_class].backward()        weights = torch.mean(self.grads, dim=[2, 3], keepdim=True)        cam = F.relu(torch.sum(weights * self.feature_maps, dim=1)).squeeze().cpu().numpy()        return camdef preprocess_image(image_path):    mean = [0.485, 0.456, 0.406]    std = [0.229, 0.224, 0.225]    transform = transforms.Compose([        transforms.Resize((224, 224)),        transforms.ToTensor(),        transforms.Normalize(mean, std)    ])    img = cv2.imread(image_path)    img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)    img = transform(img).unsqueeze(0)    return imgdef visualize_cam_on_image(image_path, cam):    img = cv2.imread(image_path)    img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)    cam = cv2.resize(cam, (img.shape[1], img.shape[0]))    cam = (cam - np.min(cam)) / (np.max(cam) - np.min(cam))    cam = np.uint8(255 * cam)    heatmap = cv2.applyColorMap(cam, cv2.COLORMAP_JET)    result = cv2.addWeighted(img, 0.5, heatmap, 0.5, 0)    return resultif __name__ == "__main__":    model = models.resnet50(pretrained=True)    target_layer = model.layer4[-1].conv3    grad_cam = GradCAM(model, target_layer)    image_path = "path/to/your/image.jpg"    input_tensor = preprocess_image(image_path)    cam = grad_cam(input_tensor)    result = visualize_cam_on_image(image_path, cam)    cv2.imwrite("grad_cam_result.jpg", cv2.cvtColor(result, cv2.COLOR_RGB2BGR))

import cv2import numpy as npimport osfrom skimage import color,ioimport imageio
def cal_label(img_path,heatmap_path):    img_raw = cv2.imread(img_path)    raws_row, raws_col, raws_channels = img_raw.shape
    # 热力图    img = imageio.imread(heatmap_path)    img = cv2.resize(img, (raws_col, raws_row))    name = os.path.basename(img_path)    min_val, max_val, min_index, max_index = cv2.minMaxLoc(img)    labels = np.zeros((raws_row,raws_col,3))    for i in range(raws_row):        for j in range(raws_col):            if img[i,j] > max_val*0.9:                labels[i,j] = [255,0,0] #蓝            elif img[i,j] > max_val*0.8:                labels[i,j] = [0,255,0] #绿            elif img[i,j] > max_val*0.7:                labels[i,j] = [0,0,255] #红            elif img[i,j] > max_val*0.6:                labels[i,j] = [255,255,0]  #青            elif img[i,j] > max_val*0.5:                labels[i,j] = [0,255,255]  #黄            elif img[i,j]> max_val*0.4:                labels[i,j] = [255,0,255]            elif img[i,j] > max_val*0.3:                labels[i,j] = [60,179,113]            elif img[i,j] >max_val*0.2:                labels[i,j] = [0,191,255]            elif img[i,j] > max_val*0.1:                labels[i,j] = [255,165,0]    #粉            else:                labels[i,j] = [0,0,0]    imageio.imwrite('./output/label/'+ name, labels)    return lables

def add_light(img_path,label,label_index = 0):    output_name = os.path.basename(img_path)    img = cv2.imread(img_path)    rows, cols = img.shape[:2]    # 设置中心点    centerX = rows / 2    centerY = cols / 2    radius = min(centerX, centerY)    # 设置光照强度    strength = 300    # 新建目标图像    dst = np.zeros((rows, cols, 3), dtype="uint8")    # 图像光照特效    for i in range(rows):        for j in range(cols):            B = img[i, j][0]            G = img[i, j][1]            R = img[i, j][2]            if label[i,j] == label_index:            # 计算当前点到光照中心距离(平面坐标系中两点之间的距离)                distance = math.pow((centerY - j), 2) + math.pow((centerX - i), 2)                if (distance < radius * radius):                    # 按照距离大小计算增强的光照值                    result = (int)(strength * (1.0 - math.sqrt(distance) / radius))                    B = img[i, j][0] + result                    G = img[i, j][1] + result                    R = img[i, j][2] + result                    # 判断边界 防止越界                    B = min(255, max(0, B))                    G = min(255, max(0, G))                    R = min(255, max(0, R))                    dst[i, j] = np.uint8((B, G, R))                else:                    dst[i, j] = np.uint8((B, G, R))            else:                    dst[i, j] = np.uint8((B, G, R))    # 显示图像    output_dir = './output/lights'    if not os.path.exists(output_dir):        os.makedirs(output_dir)    cv2.imwrite(output_dir+'/'+output_name+'_%d.jpg' %label_index, dst)     if __name__ == '__main__':    image_path = "shenzhineng.jpeg"    heatmap_path = "shenzhineng_heatmap.jpg"    labels = cal_label(image_path, heatmap_path)    if not os.path.exists('./output'):        os.makedirs('./output')    for i in range(10):        add_light(image_path, labels,i)

def GussianFunction2D(x, y):    sigma = 25 #sigma为标准差，sigma越大, 高斯函数图像越扁, 则模糊效果越明显;                #sigma越小, 趋向0的话, 则模糊效果越差越接近真实图像    return (1/(2*math.pi*sigma**2)) * math.e**(-(x**2 + y**2)/(2*sigma**2))    def cal_weight_matrix(kernel=10):     #计算高斯权重矩阵    weight_matrix = np.zeros((kernel+1,kernel+1))    pos = kernel/2    # weight_matrix = np.zeros((raws,cols))    for i in range(11):        for j in range(11):            weight_matrix[i,j] = GussianFunction2D(i-pos,j-pos)    print (weight_matrix)    weight_matrix /=weight_matrix.sum()    return weight_matrix    def gaussian_value(img,i,j,weight_matrix,kernel=10):    gau_value = []    pos = int(kernel/2)    print (pos)    temp = img[i-pos:i+pos+1,j-pos:j+pos+1,:]    #根据权重矩阵计算像素值    for k in range(3):        value = np.multiply(weight_matrix,temp[:,:,k])        res = value.sum()        gau_value.append(res)    return gau_value
def add_blur(img_path,heatmap_path,labels,label_index):    img_raw = cv2.imread(img_path)    raws_row, raws_col, raws_channels = img_raw.shape    # 热力图    img = cv2.imread(heatmap_path)    img = cv2.resize(img, (raws_col, raws_row))    output_name = os.path.basename(img_path)    kernel = 10    img_gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)    min_val, max_val, min_index, max_index = cv2.minMaxLoc(img_gray)    print(max_val, min_val)    raws, cols, channels = img.shape    weight_matrix = cal_weight_matrix()    pos = int(kernel / 2)    cal_num = 0    for i in range(pos, raws - pos):        for j in range(pos, cols - pos):            if labels[i, j] == label_index:                res = gaussian_value(img_raw, i, j, weight_matrix)                img_raw[i, j] = np.array(res)                cal_num+=1    output_dir = './output/blur/'    if not os.path.exists(output_dir):        os.makedirs(output_dir)    cv2.imwrite(output_dir+output_name+'_%d.jpg' %label_index, img_raw)
if __name__ == '__main__':    image_path = "gangao.jpeg"    heatmap_path = "gangao_heatmap.jpg"    labels = cal_label(image_path, heatmap_path)    if not os.path.exists('./output'):        os.makedirs('./output')    for i in range(10):        add_blur(image_path, heatmap_path, labels,i)