Design of Automated Visual Inspection System for Beverage Industry Production Line

Quality inspection of the finished product is a significant challenge in the production process of an industrial manufacturing plant. A multistaged inspection comprising of different parts, sub-parts, sub-assemblies, and final products is required to improve the finished product’s quality and reliability in a mass production facility. Visual inspection of the finished product quality is the industry’s norm because using visual inspection, a functional defect or a cosmetic defect can be ascertained easily. A human-based visual inspection system is slow, erratic, expensive, and less accurate. Therefore, the modern manufacturing industry focuses on automated visual inspection systems using high-end imaging devices and high-power processing platforms [1].

A review of the industrial vision system can be found in the studies [2-8]. Figure 1 presents the flow of product in an assembly line where the raw material is converted to a finished product in different steps, and after the finished product is available, a set of the inspection system is initiated to inspect the quality of the finished product. Once the quality of the finished product is ascertained, then it is packed and dispatched to the market.

Visual inspection system can have four types of inspection, such as:

• Inspection of dimensional quality
• Inspection of surface quality
• Inspection of correct assembling
• Inspection of accurate or correct operation

Designing an automated visual inspection system is complicated as it requires multiple high-resolution cameras with high-speed capturing and processing facilities. Dedicated illumination and optical system play a vital role in the automated visual inspection system. Multiple vision platforms are used for the inspection of the different features of the finished product. Every vision platform comprises multiple camera and illumination sources used for image acquisition and image processing. For image acquisition, geometric camera calibration is one of the essential steps. Camera parameters include intrinsic, extrinsic, and distortion coefficients.

1.png

Figure 1. Flow of product in an assembly line

Pinhole camera model and lens distortion model of camera calibration is widely used [9, 10]. Machine vision-based can-end inspection system has been discussed by Chen et al. [11] and vision inspection system for beer bottle has been discussed by Duan et al. [12]. Feature extraction for fill level and bottle cap inspection in the bottling machine has been proposed by Yazdi et al. [13]. Vision-based liquid particle inspection of pharmaceutical injection has been proposed by Zhang et al. [14]. A vision-based system for empty bottle inspection has been reported by Huang et al. [15]. Saliency detection and template matching techniques have been used for visual inspection of the bottom of the glass bottle [16]. Morphological analysis and machine learning have been used to detect liquid impurities [17]. The empty bottle detection technique has been discussed by Duan et al. [18]. Vision-based inspection of olive oil bottling plant has been discussed by Abdelhedi et al. [19]. A programmable logic controller (PLC) has been used to control the conveyor belt of a bottling plant [20]. A detailed discussion on industrial vision system with multiple cameras, host PC, PLC, manipulators and cameras are discussed by Martínez et al. [21].

A comprehensive survey of modularity of PLC and PLC based software for automated production system has been outlined in the study [22]. Recently deep learning based technique is used to classify bottles using machine vision approach [23, 24].

This paper provides a tutorial aspect of the design of an automated visual inspection system. The dataset used for simulation is provided by Solomon and Breckon [25]. The dataset comprises 141 images of the bottle leaving a hypothetical bottling facility. A high-speed camera has captured the images at a near-constant lighting condition. The motion blur of the image has been eliminated. A visual inspection system has been developed in software where the algorithm detects the faulty bottle. To validate the simulation result, a laboratory prototype vision inspection model has been developed and real-time image acquisition and processing has been carried out.

This paper is organized as follows. Section 2 provides the problem formulation for an automated visual inspection system. Section 3 provides the simulation results. Section 4 describes the experimental setup and results obtained from experimentation. Section V concludes the paper.

Figure 2 shows the original image database used in this study. The image database comprises of multiple images which indicate different faults in the bottle (discussed in Section 2).

2.png

Figure 2. Image database [25] (a) Underfilled, (b) Overfilled, (c) No label, (d) Faulty label, (e) Label not straight, (f) Missing Cap, (g) Deformed bottle, (h) Ideal bottle

For image processing operation, first the image is converted to greyscale and binary image. The image contrast has been enhanced using the following equation:

$g(x, y)=\left\{\begin{array}{cc}\frac{11}{27}(f(x, y)-120)+200 & 120 \leq f(x, y) \leq 255 \\ \frac{37}{18}(f(x, y)-30)+15 & 30 \leq f(x, y) \leq 120 \\ \frac{1}{2} f(x, y) & 0 \leq f(x, y) \leq 30\end{array}\right.$  (1)

where, f(x, y) is the original image and g(x, y) is the output image which is a grayscale image. The coordinates of a particular bottle in the image are computed and the compared with the coordinates of the other bottles in the image. A threshold is computed and the coordinate of the damaged bottle is computed where region of interest (ROI) is found. The image processing steps followed are simple, accurate and takes less computational time.

Figure 3 presents the output of image processing algorithm where the algorithm detects the faults. In Figure 3(a) shows the detection of overfilled bottle as well as the bottle which has improper labeling. Figure 3(b) presents the detection of a void space between two consecutive bottles. Figure 3(c) detects multiple faults such as bottle with no cap and underfilled or empty bottle with cap. Figure 3(d) detects the overfilled bottle. Figure 3(e) and Figure 3(f) detects empty or underfilled bottle and bottle with no labelling.

3.png

Figure 3. Output of vision system with fault detection at each case

Different performance metrices have been calculated for the simulation program and these are summarized in Table 1. From the performance metrices it can be ascertained that the sensitivity and precision of the image processing algorithm is high and false discovery rate is low.

Table 1. Performance metrices

where, TP stands for true positive, TN stands for true negative, FP stands for false positive and FN stands for false negative.