2.1 MVGG-19

The Visual Geometry Group (VGG) makes most of the various customary leveled multipath designs of VGG19, while it associates early and late convolutional obstructs by developing additional ways of feature map handling. In this approach, three BD layers (batch normalization, drop out, global average pooling) yielding consecutive results are connected and taken care of with a brain organization of 2500 neurons.

Fig. 2 Illustrate the architecture of MVGG19 The thought behind this adjustment is to detach early and late separated highlights from a consecutive design, and interface them in a straightforward manner with the classifier at the highest point of the organization. Along these lines, the absolute removed highlights are expanded, and every way is answerable for moving elements from the early and the late picture process. In this manner, the BD blocks are not adding some other convolution activity, yet are simply moving, normalizing, and decreasing their elements ^[1].

2.2 Motif Discovery Approach

This approach is combined with a CNN using the ResNet34 architecture. Fig. 3 shows the process flow of Motif approach.

This structure has two steps.

· Images are converted into a time series.

The images are changed to a time series by utilizing a histogram. It acquires the power appropriation of the picture where the X-pivot addresses the force esteem, and the Y-hub addresses the relating pixels that have that force esteem.

· To determine image motifs, use a time series.

In this approach, a triple succession table is prepared with the help of the variable Euc. It stores values of the row, column, Euclidean distance, then Scount; the number of images in the time series is Scount[i]. Si<r for Euclidean distances in the image time series ^[2].

2.3 EfficientNet-B0

· This approach helps to increase the accuracy with minimal computational cost. It is used for a new scaling process in neural networks called compound scaling. These processes are done at the preprocessing stage. Then, the pre-trained images are fetched into a simple convolutional neural network.

· The EfficientNet-B0 baseline network is the mobile baseline network to evaluate the scaling method in convolutional networks. Table 2 shows the EfficientNet-B0 architecture ^[3].

2.4 Photometric Stereo Approach

The photometric stereo approach right off the bat recovers surface directions and can be joined with a mixed strategy to compute a stature or profundity map. Indeed, even without a resulting reconciliation step, the surface directions can be utilized, for instance, to decide the bend boundaries of article surfaces. To obtain photometric stereo pictures, the item is continuously illuminated by several light sources, illustrated in Fig. 4.

To inspect a component with this approach, four light sources are placed at different angles. The gathered five photometric images are, a curvature image, a texture image, gradient images (X & Y), and a range image. Furthermore, these images are processed by the CNN for image classification. The processing time to detect defects in a product's surface was 707ms with a CPU.

2.5 Research Gap Identified

A literature review showed the capacity of a CNN for programmed deformity discovery. In addition, the significance of the picture quality, which is exceptionally related to obtaining a satisfactory setup, is indicated by the attributes of the surface to be investigated. Uniquely, in specular surfaces, geographic data are extremely pertinent, and hence are a significant wellspring of information on location. Moreover, handling time is basic in adapting to the review of parts in high-creation-rate situations. The following are the findings from the existing methods.

· Refinement is required to prepare a trained model to improve accuracy.

· Optimization is needed to improve the processing time of a system.

· Improvising precision in defect identification is done with real-time augmented data.

· Robustness is required to train the model to reduce the false rejection rate from the input datasets.

To provide the best solution for the findings mentioned above, we propose a CNN with a DenseNet architecture for defect detection in casting products. We demonstrate the inspection system with large input image datasets to classify defective products. Through this, manufacturing industries can ensure high quality in their products.

3.1 System Architecture

The overall process of the system architecture for casting quality visual inspection using CNN with DenseNet is shown in Fig. 5.

The process flow of the architecture is as follows.

1. Casting image datasets are imported from the local directory using the image data generator function to preprocessing the images.

2. The model is prepared with the help of a CNN architecture with two stages.

3. Then, the CNN will have a database from the best-trained model to find defects in an image object.

4. Prediction of defective and non-defective datasets can be made through the linear regression process of the system.

5. Results are produced by importing the sample image data for testing from the corresponding directory.

3.2 Data Preprocessing

The first and foremost step in this work is the preparation of data to normalize pixel values (between 0 and 255) into a range between 0 and 1, aided by passing rescale arguments in the Keras ImageDataGenerator for both training and testing sets.

From the data, 20% is for validation using the argument validation_split=0.2 in the training and testing data generator, and by using the flow_from_directory~() function during preprocessing. Fig. 6 shows the distribution ratio of the datasets.

3.3 Preparation of the Trained Model

This section elaborates on the different processes involved in preparation by the best-trained model using a CNN.

2D Convolution: This is used for filtering during image processing. Here, it is developed with a 150${\times}$150 kernel with 32 filters in one stage; in another stage, the kernel size is 75${\times}$75 with the same filter size in the 2D convolution. The following equation is used for 2D convolution to provide a feature map from preprocessed data:

where

$\begin{align*} &l-Input\,\,image;\,\,m-Kernel; \\ a& b-rows\,\,and\,\,columns\,\,of\,\,output\,\,matrix\,\,indexes \\ p& q-indexes\,\,of\,\,kernel\,\,in\,\,the\,\,convolution \end{align*}$

Max_pooling 2D: To reduce the dimensions of an image, this function is used. There are two layers in this architecture: one is 75${\times}$75-dimension reduction from the 150${\times}$150 kernel, and the other reduces the kernel size from 38${\times}$38 to 19${\times}$19.

Flatten: Used to generate one-dimensional data from 2D data.

DenseNet: Used to load the pretrained data images for defect identification.

Table 3 and Fig. 7 show detailed descriptions of each layer's kernel size.

As mentioned above, the CNN comprises various layers: a progression of convolutional layers (with enactment), pooling layers, and one final, completely associated layer that creates a bunch of class scores for a given picture. The convolutional layers of the CNN go about as component extractors; they extricate shape and shading designs from the pixels that benefit from preparing the pictures.

4.1 Simulation Results

Fig. 9 shows the precision of the model, which (for the most part) increases, while errors are reduced over time. Likewise, preparation and approval of bends are firmly adjusted, indicating that the model does not cause overfitting, and may perform well when grouping pictures from the testing dataset.

Epoch 20 provided the best performance with the following results:

99.64% training accuracy

99.40% validation accuracy

1.65% training loss

2.56% validation loss

The datasets used in this study are publicly available from Kaggle.com to prepare the model to classify OK and DEF image products using approaches by the various researchers, such as a CNN with MVGG, a CNN with the Motif Discovery approach, a CNN with EfficientNet-B0, and a CNN with the photometric approach. In this work, we propose a CNN with DenseNet to classify the images. Hence, researchers can replicate this work using a variety of variations in the future.

4.2 Evaluation Criteria

The confusion matrix represents the basic prediction results from the system, which produces the outcome from binary classification. The data instances are predicted as either positive or negative. The following predictions can be made through this confusion matrix.

1. True Positive (TP): Correct Positive Prediction

2. True Negative (TN): Correct Negative Prediction

3. False Positive (FP): Incorrect Positive Prediction

4. False Negative (FN): Incorrect Negative Prediction

Accuracy: The accuracy from the images used to predict defective items can be obtained from Eq. (2):

Precision: Precision finds all positive samples from the given datasets. It is obtained from Eq. (3):

Recall: This cannot mark negative samples from the model as positive. It is obtained from Eq. (4):

F1-score: The harmonic mean of precision is called the F1-score. It is obtained from Eq. (5):

Table 4 shows a classification report from the proposed method. The mathematical expressions used to find Precision, Recall, and F1-score are discussed below.

4.3 Performance Measures

Beyond finding the accuracy of the trained model, another critical point to investigate in the system is processing time during an inspection. Measuring the time spent by the system ensures continuous improvement in production. In this connection, our system inspects castings in less processing time (around 454ms with a CPU).

Table 5 compares Precision, Recall and F1-score from existing methods for casting defect detection against our system.