Automated image processing is always worthwhile when parts arrive at a station in a disorderly manner. This also applies when the orientation varies or the position is imprecise. In such cases, a robot must first determine where the workpiece is located. If, on the other hand, each part arrives sorted and in the correct orientation, a mechanical alternative—such as feeding technology or a simple sensor—is often sufficient. This guide explains four key points: when machine vision is essential, when it’s unnecessary, which components you need, and how much a system costs.
Automated image processing is worthwhile under three conditions: disordered or varying part orientation, varying characteristics (shape, color, position, surface), or a visual inspection requirement. If any of these apply, machine vision demonstrates its advantage. Otherwise, a mechanical solution is usually more cost-effective.
Image processing is the answer to uncertainty. As soon as a machine needs to know where a part is located or whether it’s in good condition, it requires visual information. A cobot retrieves components from a box. To do this, it needs a camera that determines the part’s position and orientation. This is precisely where machine vision replaces the human eye. Only then does automation become possible at all.
When parts are already fed in an organized manner, the equation is reversed. Where parts arrive sorted, aligned, and of consistent quality, machine vision offers little added value. A vibratory spiral conveyor or a mechanical alignment rail positions every part in the same orientation—without any cameras, software, or calibration. The rule of thumb is therefore: Automate the alignment mechanically whenever possible. Only use machine vision where this alignment cannot be achieved mechanically.
Automated machine vision is always essential when a process must handle parts with undefined orientations, varying features, or visual quality requirements. With an orderly, consistent feed, a mechanical alternative is generally more robust and cost-effective.
Industrial image processing (machine vision) is the technology by which a machine captures and evaluates images and uses them to make a decision. A vision system consists of four core components: a camera with an image sensor, a lens, lighting, and evaluation software.
Industrial image processing always follows the same sequence. The lighting creates a defined contrast. The lens focuses the workpiece sharply onto the image sensor. The camera digitizes the image. The software evaluates it. Classic pattern recognition methods (edge detection, blob analysis, template matching) are used for evaluation. For cases that fall outside fixed rules, machine learning is increasingly being incorporated.
Each component contributes to the result. A high-resolution image sensor is of little use if the lens produces a blurry image or the lighting casts shadows. The choice of camera depends on the process:
| Camera Type | Operating Principle | Typical Application |
|---|---|---|
| Area-scan camera | Captures a complete 2D image in a single shot | Presence, position, and completeness checks |
| Line scan camera | Scans moving objects line by line | Continuous material, webs, roll stock, high resolution |
| Smart camera | Camera with integrated processor and software | Compact individual inspections without an external computer |
| Vision Sensor | Preconfigured sensor for a specific task | Code reading, presence detection, simple yes/no inspection |
The software is the real game-changer. It determines whether a system simply checks a single feature or interprets a complex scene. You can find an overview of vision components and complete camera kits on the Unchained Robotics marketplace.
Automated image processing covers five core applications in production: quality control, robot guidance (bin picking and pick-and-place), identification (barcodes and 2D codes), completeness checks, and deep learning-based defect detection. They all share a common goal: real-time visual decision-making.
Image processing replaces or supplements human visual inspection. It steps in where human inspection is too slow, too expensive, or too prone to error. The most important use cases in detail:
Deep learning-based image processing is the fastest-growing application. It solves inspection tasks where rule-based pattern recognition fails because the defect cannot be clearly defined visually.
Automated image processing comes in two basic forms. 2D vision captures surface area and contours (position in X and Y, contrast, code). 3D vision additionally captures height (Z-axis) and thus shape, volume, and spatial orientation. The rule of thumb: 2D for flat features, 3D for determining grip points.
2D vision is sufficient for the vast majority of inspection tasks. Often, the relevant feature lies in a single plane—such as a code, an edge, a colored area, or the presence of an object. In such cases, a 2D area scan camera delivers fast and cost-effective results. You need 3D vision when height determines success or failure. This applies when reaching into a box, during flatness inspection, or when measuring volumes.
| Criterion | 2D Image Processing | 3D Image Processing |
|---|---|---|
| Information Captured | Area, contour, contrast, code | Additionally, height, shape, volume |
| Typical Tasks | Code reading, presence detection, dimensional errors in the plane | Bin picking, grip point, flatness, volume |
| Sensitivity to light | High (contrast is critical) | Low (geometry is decisive) |
| Computational complexity | Low, capable of real-time processing | Higher; point cloud must be processed |
| System costs | Low to medium | Medium to high |
| Integration effort | Low | Higher (calibration, grasp planning) |
In practice, both approaches are often combined. A system reads the code using 2D technology and determines the gripper point using 3D technology. It is important to base the system selection on the most difficult subtask, not on the average.
In practice, machine vision systems fail due to five factors: fluctuating ambient light, dust and contamination, transparent or shiny parts, shadows, and workpieces that are difficult to detect (dark, small, similar in shape). All five can be largely mitigated with the right lighting and optics.
Image processing stands or falls with the image. The most common cause of unstable systems is not the software, but poor or inconsistent lighting. If you don’t control the ambient light, you’ll get different results in sunlight than at night. Targeted lighting is therefore the most important control factor:
Transparent parts (glass, film, clear plastic) and highly reflective surfaces are the most challenging cases. They allow light to pass through or reflect it uncontrollably. In these cases, transmitted light, polarizing filters, or dome lighting are helpful. A sealed housing keeps dust and dirt away from the optics. The key takeaway: An image processing system is only as good as its control over imaging conditions.
Image processing is particularly worthwhile when part orientation is unpredictable. If parts arrive in an orderly manner, a fixed feeding system, an sorting system, a proximity sensor, or simple barcode reading are often more cost-effective and robust. The decision is based on four criteria: part orientation, feature variance, the need for visual inspection, and production volume.
Image processing costs more than a mechanical solution, both in terms of purchase and maintenance. It is only justified when its flexibility is needed. The following matrix classifies the alternatives:
| Solution | Suitable if | Limit |
|---|---|---|
| Image Processing | Part orientation unknown, features vary, visual inspection required | Higher costs, sensitive to light and dirt |
| Feeding technology (spiral conveyor) | Parts can be mechanically positioned in a fixed orientation | Only for a small number of consistent part types |
| Organization system (nest, rail) | Parts arrive already sorted, with fixed geometry | Does not handle chaos or variations |
| Proximity sensor | Pure presence or position at a fixed location | No information on shape or quality |
| Barcode scanner | Only identification required; code must be present | No orientation or error detection |
The decision is rarely “either/or.” Often, the most cost-effective automation is a combination: mechanical separation plus a simple vision sensor to address residual uncertainty. Only when order cannot be established mechanically does a full-fledged machine vision system become the only option. In these cases, the rule is: either automate with machine vision or not at all.
The cost of a vision system varies greatly depending on its complexity. Prices range from a few hundred euros for a vision sensor to tens of thousands of euros for a 3D bin-picking system. How quickly it pays for itself depends on the manual labor saved and the scrap avoided—not on the camera price alone.
Many people judge image processing by the price of the camera. This is misleading. The camera and software rarely account for the largest portion of the cost; instead, it’s lighting, mechanical integration, calibration, and application engineering. The following ranges are industry-standard guidelines for hardware:
| System Class | Typical Price Range | Suitable for |
|---|---|---|
| Vision Sensor | 300 to 2,000 euros | Code reading, presence detection, simple yes/no checks |
| Smart camera | 1,500 to 8,000 euros | Compact 2D inspection without an external computer |
| Full 2D system | 10,000 to 30,000 euros | Camera, optics, lighting, image processing, integration |
| 3D bin-picking system | 30,000 to 80,000 euros | Reaching into the bin, including robot guidance |
The investment usually pays for itself through savings on labor costs. If a vision system replaces a manual visual inspection in a multi-shift operation, it often pays for itself within one to two years. Added to this are savings from avoided scrap and reduced complaint-related costs. Defective parts in the field can quickly exceed the system’s price. [AUTHOR INPUT REQUIRED: Does Unchained Robotics specify a concrete price range or payback period for a turnkey vision cell (e.g., MalocherBot with camera) that can serve as a reliable benchmark here?]
The cost of the camera is the smallest part of the bill. The cost-effectiveness of a machine vision system is determined by the manual labor saved, the scrap avoided, and the integration costs—not by the hardware alone.
Image processing is what makes a robot or cobot truly flexible. Instead of fixed, pre-programmed positions, the system reaches for the part exactly where it is located. This technology is called vision-guided robotics and requires three components: a camera, hand-eye calibration, and gripper planning.
Robots without vision operate blindly. They move to fixed coordinates and assume that every part is located exactly there. Only the camera breaks this rigid constraint. Hand-eye calibration translates the camera’s image coordinates into the robot’s coordinate system. This allows the Tool Center Point (TCP)—the center of the tool on the gripper—to move precisely to the detected grasping point.
The two most common scenarios:
For this interaction to work, the camera, gripper, and robot must be coordinated with one another. The article on selecting end effectors explains which end effector is suitable for the process. And if you’re just getting started, the overview of cobots and their applications provides the foundation. You can find compatible collaborative robots and cameras bundled together on the Unchained Robotics marketplace.
Automated image processing (machine vision) is the technology by which a machine captures images, evaluates them using software, and makes a decision based on that evaluation. A system consisting of a camera, lens, lighting, and analysis software detects features such as position, shape, color, or codes. It uses this information to control processes such as quality control, sorting, or guiding a robot.
Image processing is useful when parts arrive in a random order, their characteristics vary, or a process requires a visual inspection. As soon as a machine needs to know where a part is located or whether it is defect-free, only a camera can provide this information. If, on the other hand, parts arrive sorted and in the correct orientation, a mechanical alternative is usually sufficient.
Suitable alternatives to image processing include feeding technology (spiral conveyors), mechanical sorting systems (nesters, rails), proximity sensors for simple presence detection, and barcode scanners for identification. These solutions are more cost-effective and robust. However, they only work if the parts can be arranged mechanically and no visual defect detection is required.
The cost of an industrial machine vision system varies depending on the class. A vision sensor costs between 300 and 2,000 euros. A smart camera costs between 1,500 and 8,000 euros. A complete 2D system, including integration, costs between 10,000 and 30,000 euros. A 3D bin-picking system with robot guidance costs between 30,000 and 80,000 euros. Lighting and integration often account for the largest share of the cost, not the camera.
2D image processing captures the surface and contours of an object in a single plane—that is, its position, contrast, and codes. 3D image processing additionally captures height and thus shape, volume, and spatial orientation. 2D is sufficient for code reading and flat-surface inspections. You need 3D when a robot must determine a gripping point in a haphazardly filled box.
An image processing system requires four core components: a camera with an image sensor, a lens, lighting, and evaluation software. The lighting creates contrast. The lens produces a sharp image. The image sensor digitizes the image. The software analyzes it using pattern recognition or machine learning. Depending on the task, filters, enclosures, or a 3D camera may also be required.
Depending on the system, image processing operates in real time. Vision sensors and 2D systems analyze images within a few milliseconds. This allows them to keep pace with high cycle rates in production. 3D systems require more processing time because they process a point cloud. For bin picking, however, they still operate in sync with the process.
Industrial image processing primarily covers quality control, robot guidance (bin picking and pick-and-place), identification via barcodes and 2D codes, completeness checks, and deep learning-based defect detection. All of these applications share a common goal: to make a visual decision quickly and with repeatable accuracy. This applies in situations where visual inspection by humans is too slow or too prone to errors.