基于OpenCV的红外图像箱体轮廓鲁棒检测技术咨询

我需要检测红外相机拍摄图像中箱体及箱内高温物体的轮廓，目前使用classes=2的multiotsu阈值可轻松检测箱内高温物体轮廓，但无法可靠获取箱体轮廓。尝试过Canny边缘检测结合膨胀操作，效果不理想。请问是否有鲁棒的方法检测此类图像中的方形边缘？

关键约束： #

• 图像亮度差异大，但箱体始终远冷于内部物体；
• 图像为3通道红外图；
• 可接受近似箱体轮廓（无需完全匹配倾斜边）。

我曾尝试通过蓝通道直方图算法识别箱体上下边界，但仅在特定亮度条件下有效，超出条件则失效，急需更鲁棒的箱体边界检测方法。

附尝试过的代码片段：

代码片段1：Canny边缘+方形轮廓筛选

def is_square(cnt, eps=0.05, aspect_ratio_range=(0.9, 1.1)):
    # Approximate the contour to reduce the number of points
    peri = cv2.arcLength(cnt, True)
    approx = cv2.approxPolyDP(cnt, eps * peri, True)
    
    # The contour is considered a square if it has 4 vertices and is convex
    if len(approx) == 4 and cv2.isContourConvex(approx):
        _, _, w, h = cv2.boundingRect(approx)
        aspect_ratio = w / float(h)
        return aspect_ratio_range[0] <= aspect_ratio <= aspect_ratio_range[1]
    return False
blurred = cv2.GaussianBlur(img, (5, 5), 0)
edges = cv2.Canny(blurred_cp, 134,0, apertureSize=3)

vertical_kernel = cv2.getStructuringElement(cv2.MORPH_RECT, (1, 3))
dilated_vertical = cv2.dilate(edges, vertical_kernel, iterations=2)

horizontal_kernel = cv2.getStructuringElement(cv2.MORPH_RECT, (3, 1))
dilated = cv2.dilate(edges, horizontal_kernel, iterations=2)

contours, _ = cv2.findContours(edges, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
square_contours=[cnt for cnt in contours if is_square(cnt,0.05)]

cnt=square_cnt[1] #Here I tried every possible contour that i find
x, y, w, h = cv2.boundingRect(max_cnt)
cv2.rectangle(img, (x, y), (x+w, y+h), (0, 255, 0), 2)  # Draw rectangle

plt.imshow(cv2.cvtColor(img, cv2.COLOR_BGR2RGB))

代码片段2：蓝通道直方图边界检测

r,g,b= cv2.split(im)
b=cv2.bitwise_not(b)
h=cv2.calcHist([b],[0],None,[256],[0,256])
h=h[15:]
# Calculate the indices of the maxima of the histogram
indices = argrelextrema(h, np.greater)
# Get the values of the maxima of the histogram
maxima = h[indices]

# Find the indices of the 2 highest maxima
top_2_indices = maxima.argsort()[-2:][::-1]

# These are the VALUES of h at indices
maxima = h[indices]
# These are the indices of the highest values in the maxima array
top_2_indices = maxima.argsort()[-2:][::-1]

# These are the indices in h of the two highest peaks just found
highest_values = indices[0][top_2_indices]
# These are the same values, still in h, but sorted in ascending order, so the left peak appears first, and then the right peak.
highest_values_sorted=sorted(highest_values)

values_h_at_peaks= h[highest_values_sorted]
small_peak_value=min(values_h_at_peaks)[0]

_, tresh = cv2.threshold(b, 0, 255, cv2.THRESH_BINARY + cv2.THRESH_OTSU) # thresholding
tresh_padded = np.pad(tresh, ((0, 2), (0, 0)), 'constant')

row_sums = np.sum(tresh_padded, axis=1)
window_size = 3
window_sums = np.convolve(row_sums, np.ones(window_size), 'valid')

peaks, properties = find_peaks(window_sums, height=small_peak_value/2,distance=30,prominence=0.5)
peak_heights = properties['peak_heights']

top_2_indices = peak_heights.argsort()[-2:][::-1]

# Get the 3 best peaks
best_peaks = peaks[top_2_indices]

cv2.line(im, (0, best_peaks[1]), (im.shape[1], best_peaks[1]), (0, 255, 0), 2)
cv2.line(im, (0, best_peaks[0]), (im.shape[1], best_peaks[0]), (0, 255, 0), 2)

同时提供了原始图像、检测到的轮廓、Canny边缘图、期望边界框及多组不同亮度的示例图像。

内容的提问来源于stack exchange，提问作者user15233037