Basler相机强光区域泛红问题排查及代码修改咨询

• 现象：Basler相机拍摄画面中强光区域出现泛红（红色色域覆盖），其余区域显示正常，需消除该现象
• 怀疑点：代码中第40-41行的cv2.COLOR_BGR2RGB颜色空间转换及归一化操作可能导致问题，但不清楚修改方向

相关代码 #

def run_camera():        
        converter = pylon.ImageFormatConverter()
        converter.OutputPixelFormat = pylon.PixelType_BGR8packed
        converter.OutputBitAlignment = pylon.OutputBitAlignment_MsbAligned

        top_camera.OutputQueueSize = 2
        top_camera.StartGrabbing(pylon.GrabStrategy_LatestImages)

        try:
            while (True):
                top_grabResult = top_camera.RetrieveResult(5000, pylon.TimeoutHandling_Return)
                while not top_grabResult.GrabSucceeded():
                    logger.info("top grab failed, trying again")
                    top_grabResult = top_camera.RetrieveResult(5000, pylon.TimeoutHandling_Return)

                if top_grabResult.GrabSucceeded():
                    image = converter.Convert(top_grabResult)
                    package_detection_img = image.Array
                    package_detection_img = cv2.rotate(package_detection_img, cv2.ROTATE_90_COUNTERCLOCKWISE)

                    # Get the original camera resolution
                    original_h, original_w = package_detection_img.shape[:2]
                    # print(f"Current camera resolution: {original_w} x {original_h}")

                    # Define the desired size as per your original image shape
                    desired_h, desired_w = 1088, 720
                    left = 623
                    right = desired_w - left

                    # Adjusted cropping
                    cropped_width = original_w - (left + right)

                    package_detection_img = package_detection_img[:, left:left + cropped_width]

                    # Resize to desired size (original training image size)
                    package_detection_img = cv2.resize(package_detection_img, (desired_w, desired_h))
                    cv2.imshow('Live Stream', package_detection_img)  # display the window with the name 'Live Stream'
                    cv2.waitKey(1)

                    # Adjusted for brand_name_img
                    brand_name_desired_w = 108.5
                    brand_name_desired_h = 519.5
                    brand_name_crop_w = int((desired_w - brand_name_desired_w) / 2)
                    brand_name_crop_h = int((desired_h - brand_name_desired_h) / 2)
                    brand_name_img = package_detection_img[brand_name_crop_h:-brand_name_crop_h,
                                     brand_name_crop_w:-brand_name_crop_w]
                    brand_name_img = cv2.resize(brand_name_img, (int(brand_name_desired_w), int(brand_name_desired_h)))

                    # Call letterbox here, before the color conversion and normalization
                    package_detection_img = letterbox(package_detection_img, new_shape=(desired_h, desired_w))[0]
                    package_detection_img = cv2.cvtColor(package_detection_img, cv2.COLOR_BGR2RGB)
                    package_detection_img = package_detection_img / 255.0
                    # cv2.imshow('Live Stream', package_detection_img)  # display the window with the name 'Live Stream'
                    # cv2.waitKey(1)


                    # Call transformations here for both package and brand images
                    # Convert numpy arrays to PIL images for transformations
                    package_detection_img = Image.fromarray((package_detection_img * 255).astype(np.uint8))
                    brand_name_img = Image.fromarray((brand_name_img * 255).astype(np.uint8))

                    yield package_detection_img, brand_name_img

                else:
                    logger.info(f"Top camera failed to grab an image")
                    yield np.zeros((desired_h, desired_w, 3))

        finally:
            top_camera.Close()

def detect():
    if source == '1':
        for batch in dataloader:
            # print("Dataloader successfully incorporated")

            # Unnormalize before displaying
            batch = unnormalize(batch)

            frame = batch[0].squeeze(0).permute(1, 2, 0).cpu().numpy()  # Convert back to HWC format for OpenCV

            # Convert back to 0-255 range and to uint8 for OpenCV
            frame = (frame * 255).astype(np.uint8)

            # Keep a copy of the frame for display purposes
            frame_img = frame.copy()

            # Resize, convert to RGB, normalize, and convert to tensor for inference
            img = cv2.resize(frame, (int(imgsz[1].item()), int(imgsz[0])))
            img = img / 255.0
            img = np.ascontiguousarray(img.transpose((2, 0, 1)))  # HWC to CHW
            img = torch.from_numpy(img).float().unsqueeze(0).to(device)

            print(f"Transformed img shape: {img.shape}")

内容的提问来源于stack exchange，提问作者kora_num14