形态学操作

Morphology is what you do to a binary image after thresholding, and it's the most underrated cleanup step in the whole pipeline. FCV_ (menu label 形态学操作, "morphological operations") is FlowCV's Swiss Army knife for it: seven operations in one node, plus control over the kernel shape and how hard it hits. If you've ever thresholded something and gotten a mask full of speckles and broken outlines, this node is the eraser and the glue.

The pack's category for it - openCV/形态学操作 - is a hint that the author treats it as the companion to the binarization nodes. The typical flow: threshold to black-and-white, then use this node to kill the noise and close the gaps before you feed the mask to FCV_FindRectangles or a diffusion workflow.

The operations

A dropdown named 操作类型 (operation type) picks the behavior, and the tooltips are honest descriptions worth trusting:

• 腐蚀 (erode) - shrinks bright regions, removes small specks, breaks apart barely-connected objects.
• 膨胀 (dilate) - grows bright regions, fills small holes, reconnects broken bits.
• 开运算 (opening) - erode then dilate. Removes small objects while keeping big ones roughly their original size. The default, and the one you'll use most for noise cleanup.
• 闭运算 (closing) - dilate then erode. Fills small holes inside objects. The other workhorse.
• 形态学梯度 (morphological gradient) - dilate minus erode, which extracts the outline of objects.
• 顶帽 (top hat) - original minus opening; pulls out small bright details.
• 黑帽 (black hat) - closing minus original; pulls out small dark details.

Under the hood each one is a single OpenCV call - cv2.erode, cv2.dilate, or cv2.morphologyEx with MORPH_OPEN / MORPH_CLOSE / MORPH_GRADIENT / MORPH_TOPHAT / MORPH_BLACKHAT.

The knobs that matter

• 核形状 (kernel shape) - 矩形 is the default and the strongest; 椭圆 gives rounder, gentler results on circular objects; 十字 (cross) affects horizontal/vertical structures, useful when you want to preserve thin lines along one axis.
• 核大小 (kernel size, default 5) - the size of the structuring element. Must be odd; the node silently adds one if you feed it an even number.
• 迭代次数 (iterations, default 1) - how many passes. This is the "strength" dial; crank it instead of making the kernel huge, since bigger kernels get expensive fast.

Wiring and output

Input CVIMAGE in, processed CVIMAGE out - the pack's standard BGR numpy format. Since morphology assumes a binary-ish input for best results, run a threshold (fixed, adaptive, or OTSU) first, then route this output to FCV_CVToIMAGE to view it or on to the next processing step.

Installing

Bundled in FlowCV, so one install covers the whole pack. ComfyUI Manager, search "FlowCV"; or:

cd ComfyUI/custom_nodes
git clone https://github.com/Koren-cy/FlowCV

Restart ComfyUI. Dependencies are light - opencv-python, numpy, pyserial, no models. Note: the README says the project has migrated to ComfyUI_For_Academic, so this repo is in archive mode - it works, it just isn't being developed.

Gotchas

Morphology works on the whole image, so a bright object adjacent to another bright object will bleed when you dilate aggressively - that's normal, that's the operation. And remember the pack-wide silent-failure habit: on any exception the node prints a Chinese error to the console and returns your input unchanged. If the output looks untouched, look at the terminal.