Machine learning 即使没有任何对象，Resnet也显示出错误的预测

我试着用resnet对有缺陷的钢板和无缺陷的钢板进行分类。我有一个分别为这两个类别的2000幅图像的数据集。但当我对它进行实时测试时，它显示“有缺陷”，即使相机下面没有物体（无论是普通的还是有缺陷的平板）。 我在下面提供我的培训代码和测试代码。请试着找出问题所在

测试代码：

    image =frame
    #print(image.shape)
    gray = cv2.cvtColor(frame, cv2.COLOR_BGR2GRAY)
    ret,thresh = cv2.threshold(gray,160,255,0)
    contours, _ = cv2.findContours(thresh, cv2.RETR_TREE, cv2.CHAIN_APPROX_SIMPLE)
    height, width, channels = image.shape
    min_x, min_y = width, height
    max_x = max_y = 0
    for cnt in contours:
        area = cv2.contourArea(cnt)
        if area > 400:
            (x, y, w, h) = cv2.boundingRect(cnt)
            min_x, max_x = min(x, min_x), max(x + w, max_x)
            min_y, max_y = min(y, min_y), max(y + h, max_y)
            if (w>200) and (h>200):
                roi1 = image[y:y + h, x:x + w]

                roi = cv2.resize(roi1, (64, 64))
                roi = roi.astype("float") / 255.0
    # order channel dimensions (channels-first or channels-last)
    # depending on our Keras backend, then add a batch dimension to
    # the image
                roi = img_to_array(roi)
                roi = np.expand_dims(roi, axis=0)
    # make predictions on the input image
                pred = model.predict(roi)
                if (pred[0][1]) > .96:
                    label = "Defect"
                    color = (0, 0, 255)
                    confidence = pred[0][1]

                elif (pred[0][0]) > .32:
                    label = "Fair"
                    color = (0, 255, 0)
                    confidence = pred[0][0]             
                #pred = pred.argmax(axis=1)[0]

                # label = "Fair" if pred == 0 else "Defected"
                # color = (0, 255, 0) if pred == 0 else (0, 0, 255)

                cv2.putText(roi1, label, (3, 20), cv2.FONT_HERSHEY_SIMPLEX, 0.5,
                    color, 2)
                cv2.putText(roi1, "confidence:"+str(confidence*100)+"%", (3, 40), 
                    cv2.FONT_HERSHEY_SIMPLEX, 0.5,
                    (255,255,255), 2)

培训代码：

NUM_EPOCHS = 500
BS = 32

# derive the path to the directories containing the training,
# validation, and testing splits, respectively
TRAIN_PATH = os.path.sep.join([args["dataset"], "training"])
VAL_PATH = os.path.sep.join([args["dataset"], "validation"])
TEST_PATH = os.path.sep.join([args["dataset"], "testing"])

# determine the total number of image paths in training, validation,
# and testing directories
totalTrain = len(list(paths.list_images(TRAIN_PATH)))
totalVal = len(list(paths.list_images(VAL_PATH)))
totalTest = len(list(paths.list_images(TEST_PATH)))

# initialize the training training data augmentation object
trainAug = ImageDataGenerator(
    rescale=1 / 255.0,
    rotation_range=20,
    zoom_range=0.05,
    width_shift_range=0.05,
    height_shift_range=0.05,
    shear_range=0.05,
    horizontal_flip=True,
    fill_mode="nearest")

# initialize the validation (and testing) data augmentation object
valAug = ImageDataGenerator(rescale=1 / 255.0)

# initialize the training generator
trainGen = trainAug.flow_from_directory(
    TRAIN_PATH,
    class_mode="categorical",
    target_size=(64, 64),
    color_mode="rgb",
    shuffle=True,
    batch_size=32)

# initialize the validation generator
valGen = valAug.flow_from_directory(
    VAL_PATH,
    class_mode="categorical",
    target_size=(64, 64),
    color_mode="rgb",
    shuffle=False,
    batch_size=BS)

# initialize the testing generator
testGen = valAug.flow_from_directory(
    TEST_PATH,
    class_mode="categorical",
    target_size=(64, 64),
    color_mode="rgb",
    shuffle=False,
    batch_size=BS)

# initialize our Keras implementation of ResNet model and compile it
model = ResNet.build(64, 64, 3, 2, (2, 2, 3),
    (32, 64, 128, 256), reg=0.0005)
opt = SGD(lr=1e-1, momentum=0.9, decay=1e-1 / NUM_EPOCHS)
model.compile(loss="binary_crossentropy", optimizer=opt,
    metrics=["accuracy"])

# train our Keras model
H = model.fit_generator(
    trainGen,
    steps_per_epoch=totalTrain // BS,
    validation_data=valGen,
    validation_steps=totalVal // BS,
    epochs=NUM_EPOCHS)

# reset the testing generator and then use our trained model to
# make predictions on the data
print("[INFO] evaluating network...")
testGen.reset()
predIdxs = model.predict_generator(testGen,
    steps=(totalTest // BS) + 1)

# for each image in the testing set we need to find the index of the
# label with corresponding largest predicted probability
predIdxs = np.argmax(predIdxs, axis=1)

# show a nicely formatted classification report
print(classification_report(testGen.classes, predIdxs,
    target_names=testGen.class_indices.keys()))

# save the network to disk
print("[INFO] serializing network to '{}'...".format(args["model"]))
model.save(args["model"])

---

当我通过视频进行测试时，它显示出对缺陷钢板的亲和力，普通钢板的信心分数非常低，而缺陷钢板的信心分数非常高。

这样一个网络的输出是一个跨班级的概率。这不包括“nothing”类，除非您这样做。如果你在一个实验室环境中（听起来是这样），那么也许你的测试区域的规律性允许这样做。网络将始终具有一些得分最高的输出类，即使它们都非常低，尽管使用softmax激活时，类之间的总和将始终等于一。在各种情况下，有一些方法可以处理这一问题，但在很大程度上，这被认为是一个迫切需要积极研究的领域

---

除此之外，您的网络性能无法预测；这就是人工智能应用的整个领域。你需要在实验上加以改进。

那么，你建议使用二进制交叉熵或分类，以及任何其他激活函数或其他东西。你可以尝试使用sigmoid输出激活函数的二进制交叉熵损失，并获得每个类的概率区间[0,1]上的值，而不是求和到一（与softmax一样）。然而，我的想法是，如果你在受控环境下使用它，你可以用彩色胶带在你的相机下面的桌子上贴上一个大大的粉红色“X”，当你检查一个盘子时，最好将它挡住。然后你可以把粉红色的大“X”训练成“什么都没有”。谢谢你的建议，让我试试，我会把结果反馈给你