Predictive Capability Evaluation of Micrograph-Driven Deep Learning for Ti6Al4V Alloy Tensile Strength Under Varied Preprocessing Strategies

Abstract

The purpose of this study is to develop a micrograph-driven model for Ti6Al4V mechanical property prediction through integrated image preprocessing and deep learning, reducing the reliance on manually extracted features and process parameters. This paper systematically evaluates the capability of a CNN model using preprocessed micrographs to predict Ti6Al4V alloy ultimate tensile strength (UTS), while analyzing how different preprocessing combinations influence model performance. A total of 180 micrographs were selected from published literature to construct the dataset. After applying image standardization (grayscale transformation, resizing, and normalization) and image enhancement, a pre-trained ResNet34 model was employed with transfer learning to conduct strength grade classification (low, medium, high) and UTS regression. The results demonstrated that on highly heterogeneous micrograph datasets, the model exhibited moderate classification capability (maximum accuracy = 65.60% ± 1.22%) but negligible UTS regression capability (highest R² = 0.163 ± 0.020). Fine-tuning on subsets with consistent forming processes improved regression performance (highest R² = 0.360 ± 1.47 × 10⁻⁵), outperforming traditional predictive models (highest R² = 0.148). The classification model was insensitive to normalization methods, while min–max normalization with center-cropping showed optimal standardization for regression (R² = 0.111 ± 0.017). Gamma correction maximized classification accuracy, whereas histogram equalization achieved the highest improvement for regression.

1. Introduction

Ti6Al4V alloy is widely used in biomedical and aerospace applications for its excellent biocompatibility and mechanical properties [1,2]. With the development of manufacturing processes for Ti6Al4V alloy, optimizing its mechanical properties under various processing conditions has emerged as a critical research focus [3,4]. Traditionally, metallographic analyses and tensile experiments are conducted to establish the correlation between microstructural evolution and mechanical properties. However, due to their nonlinear and highly interactive interrelations, many studies remain confined to qualitative descriptions of this correlation, lacking the capability to achieve precise quantitative calculations [5].

As an artificial intelligence technology, deep learning [6] provides a novel approach to the optimization of mechanical properties in Ti6Al4V alloy, leveraging the capability to identify precise correlations within complex nonlinear systems. Many studies have employed deep learning models to predict the mechanical properties of the alloy. For example, Wang et al. [7] employed a multilayer perceptron (MLP) model, using LPBF process parameters and microstructural features as inputs, to predict the ultimate tensile strength (UTS) of Ti6Al4V alloy with a high prediction accuracy (R² = 0.907). Yang et al. [8] predicted the mechanical properties of Ti6Al4V alloy via an artificial neural network (ANN) model, integrating LPBF and HIP parameters from the literature. The model demonstrated high accuracy in predicting yield strength (YS) and UTS, with 87.5% of YS predictions and 100% of UTS predictions falling within a 5% error across all datasets. Shen et al. [9] combined LPBF parameters with defect features through feature engineering, demonstrating that MLP outperformed Bayesian networks (BN) in predicting Ti6Al4V alloy properties, achieving superior R² values of 0.954 for density (0.629 for BN) and 0.8676 for fatigue life. These studies demonstrate that deep learning models, driven by process parameters and microstructural features, can characterize mechanical properties with higher precision than traditional methods. However, such parameters-to-properties prediction methods still rely on manual analysis of micrographs for feature extraction, and the predictive performance may be compromised when faced with incomplete process parameters.

Convolutional Neural Network (CNN), a vital application of deep learning in image recognition, directly establishes correlations with targets by extracting multi-layer features from images [6,10,11]. Without relying on various process parameters, CNN models taking micrographs as input can automatically recognize microstructural features to correlate with mechanical properties. Murakami et al. [12] found that using micrographs as input, a CNN model can achieve highly accurate predictions of aluminum alloy mechanical properties. When further integrated with Grad-CAM heatmaps, the influence of different microstructural compositions on mechanical properties can be visualized. Pei et al. [13] demonstrated that image preprocessing techniques serve as a critical tool for boosting CNN model performance, as they enhanced prediction accuracy for mechanical properties corresponding to micrographs with large differences. Nevertheless, the integration of image preprocessing techniques with CNN-based prediction remains unexplored for Ti6Al4V alloy micrographs.

In this study, Ti6Al4V alloy micrographs will be collected from the published literature. A preprocessing pipeline incorporating standardization strategies and image enhancement methods will be developed. Subsequently, the CNN-based deep learning model will be employed to qualitatively classify strength grades from micrographs, providing a preliminary assessment of model performance. Regression tasks will then be conducted using the CNN-based model and compared with traditional predictive models to systematically evaluate the predictive capability. Finally, the impact of different image preprocessing combinations on model performance will be analyzed.

2. Materials and Methods

The dataset construction, preprocessing pipeline design, and deep learning model implementation in this study follow the workflow illustrated in Figure 1.

2.1. Data Collection

To avoid overwhelming structural differences, 180 micrographs with similar magnifications were selected to construct the dataset, and the size distribution of each image is detailed in Figure 2a. The corresponding ultimate tensile strength (UTS) data ranged from 850 MPa to 1250 MPa, and Figure 2b presents the specific distribution. The left scatterplot links each micrograph to its specific UTS value, and the histogram (right) shows the frequency distribution of micrographs within defined UTS intervals. The micrographs included traditional manufacturing (e.g., casting, rolling) and additive manufacturing [e.g., Selective Laser Melting (SLM), Electron Beam Melting (EBM), Directed Energy Deposition (DED)] optical micrographs, supplemented with scanning electron microscope (SEM) images for diversity enhancement.

2.2. Image Standardization

2.2.1. Color Processing

As the micrographs were collected from the published literature, Photoshop software (Adobe Photoshop 2024, version 25.0.0, Adobe Inc., San Jose, CA, USA) was used to remove the text and rulers labeled on the images. Furthermore, the inconsistent illumination conditions led to significant color differences among images. A comparative study [13] found that maintaining the original color of the material micrographs does not significantly enhance the performance of the deep learning model. Therefore, to ensure consistency among micrographs, grayscale transformation was applied to all images before model training. The processing effect is shown in Figure 3.

2.2.2. Size Adjustment

During model training, image sizes need to be kept consistent. However, the images in the dataset varied in size. Therefore, center cropping and bilinear interpolation were independently adopted to resize them to 224 × 224.

• Center cropping

Center cropping resizes an image by cropping its central region. For images smaller than the target size, zero padding is first applied to expand the image, followed by cropping to the target size. Hashemi [14] pointed out that padding with zeros does not reduce the training effect and also improves training speed of the model. The effects of center cropping are shown in Figure 4 and Figure 5.

• Bilinear interpolation

Bilinear interpolation adjusts the size by scaling the image. When the values and coordinates of the four neighboring pixels are known, bilinear interpolation computes the pixel value at the target position by performing linear interpolation in both horizontal and vertical directions [15]. Assume the scaled pixel corresponds to the original image coordinate (x, y), with four surrounding pixel values denoted as P₁₁, P₁₂, P₂₁, P₂₂, located at integer coordinates (x₁, y₁), (x₁, y₂), (x₂, y₁), and (x₂, y₂), where x₂ = x₁ + 1 and y₂ = y₁ + 1. Bilinear interpolation first performs linear interpolation along the x-axis to calculate intermediate values P₁ and P₂, as detailed in Equations (1) and (2). The final interpolated pixel value P is then derived through linear interpolation along the y-axis using P₁ and P₂, as detailed in Equation (3). The effect after scaling is shown in Figure 6.

$$P_1={\displaystyle \frac{x_2-x}{x_2-x_1}}{P}_{11}+{\displaystyle \frac{x-x_1}{x_2-x_1}}{P}_{21}$$

(1)

$$P_2={\displaystyle \frac{x_2-x}{x_2-x_1}}{P}_{12}+{\displaystyle \frac{x-x_1}{x_2-x_1}}{P}_{22}$$

(2)

$$P={\displaystyle \frac{y_2-y}{y_2-y_1}}{P}_1+{\displaystyle \frac{y-y_1}{y_2-y_1}}{P}_2$$

(3)

2.2.3. Image Normalization

Image normalization is commonly used to mitigate potential issues of gradient vanishing or explosion, as well as to accelerate model convergence [16]. Distinct normalization methods adjust the distribution of data features in various ways, which may lead to different training outcomes [17]. After applying grayscale and image resizing, three normalization methods were applied to the micrographs dataset independently: linear scaling, min–max normalization, and Z-score normalization.

• Linear scaling

As the pixel value of grayscale images ranges from [0, 255], a common normalization method is used to divide each pixel value by 255 to linearly scale the data into the [0, 1] range. The linear scaling operation is detailed in Equation (4):

$$P_{norm}={\displaystyle \frac{P_i}{255}}$$

(4)

where P_norm denotes the pixel value after normalization, while P_i denotes the original pixel value.

2.

Min–max normalization

Min–max normalization processes each image according to Equation (5):

$$P_{norm}={\displaystyle \frac{P_i-\min (P)}{\max (P)-\min (P)}}$$

(5)

where min(P) and max(P) represent the minimum and maximum pixel values in the image. When min(P) = 0 and max(P) = 255, then min–max normalization simplifies to Equation (4).

3.

Z-score normalization

Z-score normalization processes each image according to Equation (6):

$$P_{norm}={\displaystyle \frac{P_i-\mu }{\sigma }}$$

(6)

where $\mu$ denotes the mean of the dataset, and $\sigma$ denotes the standard deviation of the dataset. Z-score normalization computes the mean and standard deviation for each channel across all images in the dataset to perform normalization. The processing effects and corresponding pixel value distributions are shown in Figure 7.

Compared to the original grayscale image, both linear scaling and min–max normalization resulted in minimal visual changes. However, these two methods scaled the pixel value range to [0, 1], and min–max normalization further enhanced uniformity in pixel value distribution. Z-score normalization led to a significant visual shift because the pixel values were scaled to the range [−3, 3], which is outside the displayable range of [0, 1] for OpenCV. Nevertheless, the information stored in the image was not affected by this transformation.

2.3. Image Enhancement

Image enhancement aims to improve image quality by enhancing details and optimizing visual effects. Common methods of image enhancement include contrast enhancement [18], brightness adjustment [19], and other approaches. Following the normalization, a series of image enhancement methods were applied to further evaluate the performance of the model.

Table 1. Roles and parameters of each image enhancement method.

Method	Role	Parameters
Mean Filtering (MF)	Remove noise and smooth image	Ksize = (3,3) (Based on OpenCV)
Gaussian Filtering (GF)	Remove noise and smooth image	Ksize = (3,3), sigmaX = 0 (Based on OpenCV)
Gray Linear Transformation-1 (GLT-1)	Reduce brightness and contrast	Y = kx + b, x represents pixel value, k = 0.75, b = 0
Gray Linear Transformation-2 (GLT-2)	Increase brightness and contrast	Y = kx + b, k = 1.15, b = 0
Gamma Correction-1 (GC-1)	Increase brightness and highlight dark details	Y = xγ, x is divided by 255, x ∈ [0, 1]. γ = 0.5
Gamma Correction-2 (GC-2)	Reduce brightness and highlight bright details	Y = xγ, x ∈ [0, 1]. γ = 1.7
Contrast Limited Adaptive Histogram Equalization (CLAHE)	Enhance local contrasts of the image	clipLimit = 2.0, titleGridsize = (4,4) (Based on OpenCV)
Histogram Equalization (HE)	Enhance contrasts and details of the image	Default parameters (Based on OpenCV)

lists the roles and parameters of each method. The effects of the enhancement are shown in Figure 8.

Figure 8A,B,a,b and show that even though MF and GF did not cause significant changes in visual effects, the distribution of pixel values differed from the original. This change occurred because both MF and GF removed potential noise and smoothed the image. Figure 8C,D,c,d correspond to the effects of GLT processing under two different parameter settings. The former (Figure 8C,c) decreased the brightness and contrast by reducing pixel values, while the latter (Figure 8D,d) increased them. Figure 8E,F,e,f correspond to the effects of γ = 0.5 and γ = 1.7. When γ < 1, GC nonlinearly stretched dark regions, enhancing visibility and perceived brightness. When γ > 1, GC compressed highlights, recovering details and improving contrast [20]. Figure 8G,H,g,h correspond to the effects of CLAHE and HE [21]. HE globally equalized the histogram to a near-uniform distribution to achieve contrast enhancement, while CLAHE adaptively constrains local contrast within subregions, producing smoother intensity transitions.

2.4. Models for Classification and Prediction

He et al. [22] introduced the ResNet architecture in 2015, employing residual blocks with shortcut connections to mitigate model degradation in deep networks. This innovation enabled ResNet to win the 2015 ImageNet classification competition. In this paper, a pre-trained ResNet34 model was employed for transfer learning to extract features from micrographs of Ti6Al4V alloy. As an initial step in assessing the model’s predictive capability, a classification task was performed by partitioning the dataset into three discrete strength grades based on UTS magnitude, as detailed in

Table 2. Classification of strength and corresponding sample proportion.

Strength Grade	UTS Interval (MPa)	Proportion of Samples (%)
Low-strength	$850\le \mathrm{UTS}\le 950$	28
Medium-strength	$950<\mathrm{UTS}\le 1050$	35
High-strength	$1050<\mathrm{UTS}\le 1250$	37

. To address the mild class imbalance in low-strength samples, data augmentation techniques such as random flipping (horizontal and vertical) and random rotation (degrees = 15) were applied to the training set. The structure of the classification model is illustrated in Figure 9a, where the output layer was reconfigured with three neurons corresponding to the target classes.

In the regression task, as shown in Figure 9b, the ResNet34 architecture was adapted as a CNN-based regression model by modifying its output layer to a single neuron. In addition, to compare the performance of the CNN-based model, three traditional predictive models including mean baseline, linear regression, and random forest were implemented. The mean baseline generates predictions by computing the mean of the training set labels, while both linear regression [23] and random forest [24] produce outputs through fitting image features extracted from the Gray-Level Co-occurrence Matrix (GLCM) [25].

The classification model employed cross-entropy loss function [26] as the loss function, with the calculation formula shown in Equation (7):

$$L={\displaystyle \frac{1}{n}}{\displaystyle \sum _i{\displaystyle \sum _{c=1}^M{y}_{ic}\log (p_{ic})}}$$

(7)

where n is the number of samples, M represents the number of categories, y_ic is the indicator function (1 if sample i’s true class is c, 0 otherwise), p_ic represents the predicted probability of sample i belonging to class c. The CNN-based regression model used the mean square error (MSE) [27] as the loss function, which is detailed in Equation (8).

To ensure the generalization ability of the model and to mitigate potential biases introduced by random data partitioning, the dataset was divided using 5-fold cross-validation, where each iteration allocated four subsets for training and one used for testing. Details regarding other methodologies and hyperparameter configurations are summarized in Table 3, Table 4 and Table 5, respectively.

The classification model performance was evaluated by the accuracy [28], while the regression model was evaluated by aggregating test set predictions from all cross-validation folds to compute overall coefficient of determination (R²). Furthermore, to minimize the impact of random initialization that may lead to premature convergence, 5-fold cross-validation was conducted five times for each preprocessed dataset. Then, the mean and standard deviation of accuracy and R² were computed after ruling out outliers. The calculation formula of R² is detailed in Equation (9).

$$MSE={\displaystyle \frac{1}{n}}{\displaystyle \sum _{i=1}^n{(y_i-f(x_i))}^2}$$

(8)

$$R^2=1-{\displaystyle \frac{{\displaystyle \sum _{i=1}^n{(f(x_i)-y_i)}^2}}{{\displaystyle \sum _{i=1}^n{(\frac{1}{n}{\displaystyle \sum _{i=1}^n{y}_i-y_i})}^2}}}$$

(9)

In the above equations, y_i represents the true value corresponding to the i-th image, while f(x_i) denotes its predicted value, and n is the number of samples. In this study, PyTorch (version 2.0.0, Meta Platforms, Inc., Menlo Park, CA, USA) was used as the deep learning framework.

Table 3. Methods and hyperparameter configurations for the classification model.

Method and Hyperparameter	Configuration
Optimizer	Adam
Epoch	200
Early stop patience [29]	7
Learning rate	0.0001
Batch size	8
Initialization of weights in FC	Xavier uniform
K-fold cross validation	K = 5, Shuffle = True, random_state = 42

Table 3. Methods and hyperparameter configurations for the classification model.

Method and Hyperparameter	Configuration
Optimizer	Adam
Epoch	200
Early stop patience [29]	7
Learning rate	0.0001
Batch size	8
Initialization of weights in FC	Xavier uniform
K-fold cross validation	K = 5, Shuffle = True, random_state = 42

Table 4. Methods and hyperparameter configurations for the CNN-based regression model.

Method and Hyperparameter	Configuration
Optimizer	Adam
Epoch	200
Early stop patience	16
Learning rate	0.008
Batch size	8
Initialization of weights in FC	Xavier uniform
K-fold cross validation	K = 5, Shuffle = True, random_state = 42

Table 4. Methods and hyperparameter configurations for the CNN-based regression model.

Method and Hyperparameter	Configuration
Optimizer	Adam
Epoch	200
Early stop patience	16
Learning rate	0.008
Batch size	8
Initialization of weights in FC	Xavier uniform
K-fold cross validation	K = 5, Shuffle = True, random_state = 42

Table 5. Methods and hyperparameter configurations for the traditional predictive model.

Random Forest	Configuration
	n_estimators	100
	Random state	42
GLCM	distance	[1]
	angles	[0]
	Symmetric	True
	normed	True
	levels	256
	Extracted features	contrast
		energy
		homogeneity
		correlation

Table 5. Methods and hyperparameter configurations for the traditional predictive model.

Random Forest	Configuration
	n_estimators	100
	Random state	42
GLCM	distance	[1]
	angles	[0]
	Symmetric	True
	normed	True
	levels	256
	Extracted features	contrast
		energy
		homogeneity
		correlation

3. Results

3.1. Test Results of Classification Model

3.1.1. Test Results of Image Standardization

Figure 10 presents the mean accuracy and corresponding standard deviation of the classification model under different image standardization strategies. Compared to the non-normalized data, the mean accuracy was only slightly improved on the dataset processed with interpolation and linear scaling. Other standardization strategies led to performance degradation, particularly for Z-Score normalization and min–max normalization. Based on these results, the subsequent enhancement methods were built upon the combination of linear scaling and bilinear interpolation.

3.1.2. Test Results of Image Enhancement

Figure 11 presents the mean accuracy and corresponding standard deviation of the classification model using different image enhancement methods. Compared to the accuracy in Section 3.1.1 (63.97% ± 2.36%), GC-1 achieved the highest accuracy (65.60% ± 1.22%), followed by CLAHE with comparable performance and lower variability (65.20% ± 0.71%), while GLT-2 showed minimal improvement (64.57% ± 2.63%). However, the application of HE resulted in significant degradation of model performance, demonstrating its unsuitability for the classification task.

3.2. Test Results of Regression Models

3.2.1. Test Results of Image Standardization

Table 6. Results of image standardization for CNN-based regression model.

Crop	Scale	Linear Scaling	Min–Max	Z-Score	Mean R2	Standard Deviation
√		√			0.061	0.029
√			√		0.111	0.017
√				√	0.074	0.018
	√	√			0.079	0.006
	√		√		0.061	0.031
	√			√	0.102	0.026
√					0.047	0.021
	√				0.026	0.020
Mean baseline					−0.003	0

and

Table 7. Result of image standardization for traditional predictive models.

Linear Regression	Random Forest	Crop	Scale	Mean R2
√		√		−0.024
√			√	0.003
	√	√		−0.028
	√		√	0.109
Mean baseline				−0.003

present the R² results of the CNN-based regression model and traditional predictive models on standardized datasets, respectively. Compared to the baseline result (R² = −0.003), the CNN-based model demonstrated only marginal advantages on non-normalized datasets (R² = 0.047 ± 0.021 and R² = 0.026 ± 0.020). The application of normalization methods improved the test results, and the combination of center cropping and min–max normalization yielded the highest R² value (0.111 ± 0.017). Among the linear regression and random forest models, the random forest achieved the highest R² (0.1093) on interpolated dataset. However, the inappropriate image resizing method could degrade model performance below baseline level.

These results indicated that both CNN-based models and traditional predictive models exhibited poor predictive capability, with the highest mean R² reaching only 0.111. This suggested that the models failed to precisely characterize UTS through learning microstructural features when relying only on standardized datasets. Therefore, image enhancement methods were implemented subsequently to enhance image details and further explore model predictive potential. The enhancement methods were built upon the standardization strategies corresponding to each model’s highest R² in this section.

3.2.2. Test Results of Image Enhancement

Table 8. Results of image enhancement for CNN-based regression model.

MF	GF	GLT-1	GLT2-2	GC-1	GC-2	CLAHE	HE	Mean R2	Standard Deviation
√								0.117	0.029
	√							0.126	0.017
		√						0.080	0.018
			√					0.058	0.006
				√				0.102	0.031
					√			0.089	0.026
						√		0.071	0.021
							√	0.163	0.020
Mean baseline								−0.003	0

,

Table 9. Results of image enhancement for linear regression.

MF	GF	GLT-1	GLT2-2	GC-1	GC-2	CLAHE	HE	Mean R2
√								0.093
	√							0.082
		√						0.045
			√					0.014
				√				0.052
					√			0.024
						√		−0.086
							√	−0.043
Mean baseline								−0.003

and

Table 10. Results of image enhancement for random forest.

MF	GF	GLT-1	GLT-2	GC-1	GC-2	CLAHE	HE	Mean R2
√								0.016
	√							0.085
		√						0.111
			√					0.139
				√				−0.021
					√			0.048
						√		−0.016
							√	−0.057
Mean baseline								−0.003

present the R² results of the CNN-based regression model and traditional predictive models on enhanced datasets. Regrettably, while certain enhancement methods improved test outcomes, the CNN-based models achieved the highest R² of only 0.163 ± 0.020 (HE-processed dataset), whereas the traditional models attained the highest R² of 0.139 (yielded by the random forest on GLT-2-processed dataset). As shown in Figure 12, the distribution of predicted values for the CNN-based model on the HE-processed dataset was further visualized by a scatter plot with error bars, calculated from the mean and standard deviation of multiple predictions. The figure shows that most predictions were constrained within a specific UTS interval with significant variability, demonstrating that the models had almost no practical predictive capability.

An explanation of the poor results may stem from the high heterogeneity of the dataset, which mixed micrographs from multiple forming processes. As illustrated in Figure 13, significant morphological evolution of Ti6Al4V alloy microstructures might arise due to the phase transformations driven by different processing conditions. These variations introduce high feature complexity, hindering the models’ ability to capture microstructural features that reliably correlate with UTS. Therefore, the next tests focused on creating subsets with less heterogeneity to explore model performance.

3.2.3. Test Results of Subsets with Less Heterogeneity

To reduce heterogeneity, the subsets were constructed by selecting micrographs with consistent forming processes from the dataset. Based on the previous results comparison, this section analyzed the predictive capability of the CNN-based model on HE-processed subset and the random forest on GLT-2 subset. To address limited sample sizes in subsets, the training strategy of the CNN-based model was modified as follows: A subset was isolated from the original dataset, and pre-trained weights were obtained using remaining micrographs for initial training. Subsequently, the subset was fine-tuned with reduced learning rates.

As presented in

Table 11. Results of HE-processed subsets for the CNN-based regression model.

Subset	Sample Size	Mean R2	Standard Deviation
SLM-based	56	0.298	0.003
DED-based	16	0.360	1.47 × 10−5
EBM-based	18	0.329	6.66 × 10−5

and

Table 12. Results of GLT-2-processed subsets for the random forest.

Subset	Sample Size	Mean R2
SLM-based	56	0.137
DED-based	16	0.148
EBM-based	18	-0.233

, three subsets were employed to evaluate the performance of the model. The results demonstrate that learning micrographs with less heterogeneity improved the test R² of the CNN-based model, with the highest value of 0.360 achieved after fine-tuning on the DED-based subset. Furthermore, as shown in Figure 14b,c, fine-tuning with subsets containing fewer samples significantly enhanced the model’s predictive stability. In contrast, the random forest exhibited no significant improvement, yielding a maximum R² of only 0.148.

4. Discussion

4.1. Predictive Capability of the Model

In the classification task, the model achieved a maximum accuracy of 65.60% ± 1.22%, demonstrating its moderate capability in qualitatively classifying strength grades from highly heterogeneous micrographs. However, for the regression task, both CNN-based and traditional models exhibited almost no predictive capability on the dataset (highest R² = 0.163 ± 0.020). Fine-tuning the CNN-based model on subsets could improve performance. The model fine-tuning on the DED-based subset yielded a peak R² of 0.360 ± 1.47 × 10⁻⁵. This suggested that the model was better suited in learning micrographs derived from consistent forming process, as the microstructural variations were relatively small. In contrast, the random forest showed slight improvement on subsets (highest R² = 0.148), indicating the CNN-based model has better predictive potential on datasets with less heterogeneity.

Despite these improvements, the R² values were still not ideal. This may arise from different post-processing methodologies across studies, which introduced microstructural heterogeneity even though the alloys were formed in consistent processes, as shown in Figure 15. Additionally, Tsutsui et al. [30] demonstrated that differences in microscopy instruments can induce image feature shifts, further compromising model performance. Therefore, constructing datasets with Ti6Al4V micrographs derived from similar experimental conditions may enhance predictive capability.

4.2. Impact of Size Adjustment and Normalization Methods on Model Performance

In the classification task, no normalization method improved model performance on the cropped dataset, whereas only linear scaling enhanced performance on the interpolated dataset. This suggests that the classification model was not sensitive to normalization. Conversely, normalization methods universally improved regression performance, with the combination of min–max normalization and center-cropping achieving optimal standardization. This indicates that the selection of normalization should depend on both the resizing method and the task type.

4.3. Impact of Image Enhancement Methods on Model Performance

In the classification task, GC-1, CLAHE, and GLT-2 improved model performance, with the highest accuracy achieved on the GC-1-processed dataset. These three methods all aimed to enhance image brightness and detail. However, despite HE also being implemented to enhance contrast, it failed to improve model performance. This may be because HE enhances global contrast, potentially amplifying details unrelated to strength grade classification.

For the regression task, the CNN-based model performed best on HE-processed datasets, followed by GF and MF. The improvements from MF and GF suggest the presence of noise affecting regression performance, as shown in Figure 16. GF outperformed MF, likely because GF employs weighted averaging [31], which reduces the influence of distant pixels on the center pixel to preserve image information, whereas MF relies on local averaging that may blur details. Unlike the classification model, the regression model benefited more from HE-processed datasets. This suggests that the CNN-based model tends to learn globally enhanced microstructural features, while the classification model focuses on amplified local information.

In summary, the classification model benefited from brightness enhancement and localized contrast amplification, whereas the regression model achieved optimal performance with global contrast enhancement, followed by denoising and smoothing techniques.

5. Conclusions

This study presents a systematic evaluation of the predictive capability of a CNN-based deep learning model for the UTS of Ti6Al4V alloy using preprocessed micrographs, and it analyzes the impact of different image preprocessing combinations on model performance. On highly heterogeneous datasets, the model exhibited moderate capability in quantitatively classifying strength grades from micrographs (accuracy = 65.60% ± 1.22%) but negligible UTS regression capability (highest R² = 0.163 ± 0.020). It is more appropriate to conduct the regression task using micrographs obtained under similar experimental conditions: The model fine-tuned on a subset with consistent forming process demonstrated improved performance (highest R² = 0.360 ± 1.47 × 10⁻⁵), outperforming traditional predictive models.

During standardization, the normalization selection should be matched to the specific image resizing method and task type: classification model was not sensitive to normalization, while normalization methods universally improved regression performance, with the min–max normalization and center-cropping combination achieving optimal standardization. Among image enhancement methods, gamma correction most effectively improved classification accuracy, while histogram equalization delivered optimal regression performance. The classification model performed better on datasets processed by brightness enhancement and localized contrast amplification, while the regression model benefited most from global contrast enhancement, followed by denoising and smoothing.

Overall, this research aids the researchers in selecting appropriate data preparation strategies and preprocessing approaches when applying machine learning to predict mechanical properties of Ti6Al4V alloys through micrograph analysis.