Image enhancement with art design: a visual feature approach with a CNN-transformer fusion model

Histogram equalization

HE is a technique that involves expanding areas with high pixel density to adjacent regions, creating layers, and gradually narrowing areas with fewer pixels. This method increases the overall gray level range, optimizes the utilization of different gray levels, and enhances image contrast (Zhang et al., 2021). The procedure for HE comprises the following steps:

Histogram segmentation: This step involves dividing the histogram into multiple sub-histograms, which are then processed independently. This approach aims to achieve a more natural visual effect in the enhanced image and mitigates the problem of brightness migration commonly observed in traditional HE methods.

Histogram clipping: The primary objective of this step is to regulate the enhancement rate during the HE process and to prevent the excessive stretching of certain gray levels. This is accomplished by setting thresholds on gray levels with high and low frequencies.

Independent equalization: This step involves applying the equalization process separately to each sub-histogram. The enhancement effect may be minimal for sub-histograms with narrow ranges, whereas it may be excessive for those with broader ranges. Consequently, the mapping space for each sub-histogram is readjusted to ensure a balanced dynamic range.

HE can be popularized in discrete images. If the total number of pixels in the image is assumed to be one gray level, and the value of the first gray level is assumed to have the number of gray level pixels, gray level pixels will appear in the image (Wu, Song & Zhang, 2020). If the number of pixels in an image is n, and the gray level of the picture is L, then the probability of producing a fourth gray level is m: (1)P r = n m n m = 0 , 1 , 2 , 3 , 4 … L .

The processing function after equalization: (2)F k = ∑ i = 0 m n i n .

If the equalization range is [0-255], all gray levels in the image are mapped, yielding the following function: (3)S K = 255 . F K where S(K) represents the scaled value of the pixel intensity K after equalization; 255 represents the maximum intensity value in an 8-bit grayscale image; F represents a scaling factor; K represents the original intensity value before equalization.

Bilayer column balance is an optimization of histogram and single platform balance. This method prevents excess pixels (usually including background and noise) from grayscale by setting the highest platform threshold, and prevents low pixels (usually including weak objects and details) from merging through other grayscale by setting the lower platform threshold during the enhancement process. Low pixels (usually including weak objects and details) are fused by other gray levels. The calculation formula is: (4)P k = T max P k T min P k > = T max T min < = P k < = T max 0 < = P k < = T min where the equation P(k) represents a corrected value for the dual-platform histogram, which categorizes pixels based on their intensity values. T_max represents the upper threshold for the pixel intensity. T_min: The lower threshold for the pixel intensity.

Equation (4) explains how pixel intensities are classified: P(k) ≥ T_max—pixels with intensity values greater than or equal to the upper threshold are assigned T_max; T_min ≤ P(k) ≤ T_max - pixels within the threshold range are kept as they are; P(k) ≤ T_min—pixels with intensity values less than or equal to the lower threshold are assigned T_min.

After modifying the statistical histogram, the cumulative histogram of the image is derived from the adjusted statistical histogram. The gray levels of the image are then redistributed using this cumulative histogram, akin to the HE method, resulting in an enhanced image with improved balance.

(1) Accumulate the revised histogram: (5)F T k = ∑ i = 0 k Pr k 0 < i < 25 where FT(k) is the cumulative histogram of the image.

(2) Map the gray value in the original image, taking the 8-bit image as an example: (6)S k = 255 . F T k where S_k is the statistical histogram of the image, T is the platform threshold, and FT(k) is the platform statistical histogram of the image.

AFF based CNN-Transformer dual-stream network

Figure 2 shows the network structure diagram of the first stage of the CTGAN algorithm, which adopts a U-Net type network architecture and consists of an encoder, a decoder, and a “jump part” that provides side auxiliary input for the decoding module. The encoder adopts the two-flow feature extraction path of CNN-Transformer, and realizes feature extraction from the input image step by step through multi-level feature processing layers from bottom to top. Convolution with step 2 is used between adjacent feature processing layers to perform spatial downsampling of the feature graph. The feature processing layer of each level is determined by the residual dense block (Hu et al., 2023).

The encoder sequentially encodes the input image to capture diverse feature information and furnish the decoder with varying levels of local structural features and semantic details for stepwise image reconstruction. To harness the strengths of CNN in local feature extraction and Transformer in global feature capture, this paper integrates a dual-stream feature extraction pathway at the encoder’s end, comprising CNN and Transformer branches.

The CNN branch takes the residual-intensive module RDB as the core module. Figure 3 shows the specific implementation of each RDB module. The essence of RDB module is a residual module with dense connections. Each basic block involved in dense connections is composed of 3×3 spatial convolution and ReLU operation.

Figure 3 illustrates the unique dense connectivity among basic blocks within the RDB, enabling each level’s basic block input to integrate the output feature maps from previous levels and the input feature map of the RDB block. Through this progressively intensive connection and processing, the RDB block achieves local feature extraction of input feature maps across different levels and a multiscale receptive field perception.

In this paper, the CNN feature extraction pathway is structured around a feature extraction layer centered on the RDB block. The output feature map from the deeper layers of the RDB block contains richer spatial local structural information with diverse receptive fields, built upon the subsampling of the preceding RDB block’s output feature map space. Consequently, the CNN feature extraction pathway maximizes the benefits of RDB blocks for feature extraction, facilitating bottom-up local feature extraction and gradual embedding of input images.

As shown in Fig. 4, the H-Transformer consists of hierarchical multi-head self-attention (HMHSA) (Li et al., 2023), layer normalization (LN), and a multi-layer perceptron (MLP). The H-Transformer first divides the input feature map spatially to form a sequence of subblocks, and then applies a linear transformation via Patch Embedding to generate the initial feature vectors for these subblocks. This allows the feature encoding of the input feature map to be achieved using the self-attention layer, based on the normalization of subsequent layers. The context eigenvector representation of these subblock sequences is then obtained, enabling the model to better capture long-range dependencies within the subblock sequences.

For the single HMHSA layer in Fig. 4, three MHSA sub-layers are cascated in a bottom-up sequence. If the space division process of the input feature map is expressed by a recursive quadtree with depth of 3, The lowest level MHSA sublayer of HMHSA pays more attention to the calculation of multi-head self-attention of each fine molecule block in the region where the leaf node is located, and therefore focuses on the modeling of spatial dependence between different fine molecule blocks in the smaller local region. Therefore, with the help of successive successive MHSA sublayers in the HMHSA layer, the progressive and nested self-attention calculation of the input feature map from the local region to the more global region is realized. Compared with the self-attention calculation of the entire feature map directly using the MHSA layer, the computational complexity is reduced.

To effectively combine the encoding features from CNN and Transformer branches, the attention feature fusion module AFF is introduced into each feature extraction layer of the encoder. As shown in Fig. 5, taking the information fusion of two branches of the encoder as an example, the multi-channel feature maps of the organization feature layer of the Transformer branch and the organization feature layer of the CNN branch are simultaneously added to the AFF module. In order to enhance global information from Transformer branches, channel focus enhancement and spatial focus enhancement can be carried out, which effectively enhances spatial local structure information from CNN branches.

At the same time, using 3×3 spatial convolution, the cross-path fine-grained interaction between two sets of multi-channel feature graphs is carried out. Finally, the output of AFF module is generated by channel connection and 3×3 spatial convolution. Obviously, the AFF module skillfully integrates channel attention, spatial attention and cross-path interaction mechanism, and effectively captures the global and local context information of the input image at the current feature extraction layer with the help of dual-path feature fusion, providing effective data preparation for the “jump part” between the encoder and the decoder.

GAN

To obtain high-quality enhanced images that align with human visual perception, this paper employs a conditional generative adversarial network (CGAN) in the second stage. By using unpaired high-visual-quality images from the Ava dataset (Gu et al., 2018) as input conditions, the images, which underwent initial enhancement in the first stage, undergo further colour adjustment. More detailed information is recovered through guided adversarial learning of image perceptual quality, resulting in a further enhancement of the visual effect of the initially enhanced images.

As depicted in Fig. 6, the generator adopts a U-Net architecture, which consists of an encoder, decoder, and skip connections. The encoder includes the RDB module and an undersampling layer, which is essentially a 3×3 convolutional layer with a stride of 2. The decoder comprises an RDB module and a 3×3 transposed convolutional layer with a stride of 2. In the skip connections, the low-level feature map from the RDB module at the encoder is laterally passed to the decoder, where it is added to the feature map of the same resolution generated by the decoding module, providing input for subsequent decoding.

Using the initially enhanced image from the first stage as input, the generator’s encoder extracts local features at each layer using the RDB module, gradually abstracting input image features from low to high levels. The output of the bottleneck layer is then fed into the decoder, where the feature maps from the encoder are combined through top-down skip connections for stepwise decoding. This fusion of high-level semantic information and low-level features from the initially enhanced image leads to further image enhancement.

In the second stage of the CTGAN, the enhancement process builds upon the initially enhanced image from the first stage, rather than processing the original image directly. This approach allows the generator to achieve the desired enhancement effect with a relatively shallow U-Net architecture.

Dataset and evaluation indicators

Two datasets are used in this paper, including the publicly available large-scale graphic design dataset DesignNet, linked to (https://design-net.org/, doi: https://doi.org/10.1145/3591196.3596614) and the AVA dataset, linked to (https://github.com/mtobeiyf/ava_downloader/tree/master/AVA_dataset, doi: 10.1109/CVPR.2012.6247954). The DesignNet dataset contains 4,369 visual images used in interface design, which are labeled with semantic information of style, technique, and space. Currently, each image in the dataset corresponds to a single label in a single dimension, but in interface design practice, each design image should have the same label in three different dimensions. The AVA dataset comprises over 250,000 images, each annotated with semantic labels, style labels, and aesthetic scores. It includes 66 categories of semantic labels, 14 categories of style labels, and aesthetic scores ranging from 1 to 10. To validate the proposed graphic design image enhancement model based on visual characteristics, a series of experiments were conducted. The specific experimental settings are detailed in Table 1. Table 2 outline the model parameter settings.

This paper mainly defines several evaluation indexes for image enhancement algorithms. In addition to the common P (Precision), R (Recall) and F values, there are also two indicators: peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) (Osorio et al., 2022; Prodan, Vlăsceanu & Boiangiu, 2023).

Given two images X and Y of m×n, the PSNR value of image Y relative to image X is defined as shown in Eq. (7). The unit of PSNR is dB, and the larger the value, the lower the degree of distortion of image Y relative to image X. (7)P S N R X , Y = 10 × log 10 M A X Y 2 M S E X , Y = 20 × log 10 M A X Y M S E X , Y where, MAX_Y denotes the maximum pixel value in the image Y, which is usually taken for 256-level grayscale images, MAX_Y = 255;

MSE(X, Y) is the mean square error of the pixel values of X and Y of the two images, calculated as shown in Eq. (8). (8)M S E X , Y = 1 m × n ∑ i = 0 m − 1 ∑ j = 0 n − 1 X i , j − Y i , y 2 .

SSIM combines brightness, contrast and structure to measure the similarity of two images from the perspective of image composition. Given two images of m×n X, Y, calculate the SSIM between X and Y according to Eq. (9). (9)S S I M X , Y = 2 μ X μ Y + c 1 μ X 2 + μ Y 2 + c 1 × 2 σ X Y + c 2 σ X 2 + σ Y 2 + c 2 where μ X and Y are the average brightness values of X and Y, which reflect the overall brightness of X and Y. σX and Y are the standard deviation of brightness of X and Y.

Compared with PSNR value, SSIM based image quality evaluation is more consistent with the visual characteristics of human eyes. SSIM ranges from 0 to 1. In the actual calculation, the given image pair is usually divided into spatial blocks, and the SSIM value of each image block is calculated at the same time. The SSIM value of the given image pair can be obtained by arithmetic average of the SSIM value of these image blocks.

Model evaluation methods

To eliminate noise, redundancy, and inconsistencies in the data and to enhance its standardization, reliability, and applicability for specific tasks, data preprocessing is essential. For image preprocessing, we utilize mean subtraction, which involves subtracting the mean value of all training set images from each image’s features during training. This process centralizes the data of each dimension to zero, reducing computational overhead. It transforms the data into a matrix composed of vectors under the original standard coordinate system, centered around the mean values of these vectors. This step can be achieved using Eq. (10): (10)X ′ = X − np . mean X , axis = 0 .

In this step, the mean of each feature across the entire dataset, X, is computed along the specified axis (in this case, axis 0, which refers to columns or features). The resulting mean vector is then subtracted from each data point in the dataset. This operation centers the data, ensuring that the average value of each feature is zero, which simplifies further computation and can enhance the performance of neural networks by normalizing input data. np.mean is a function from the Python library NumPy, which is used for numerical computing.

Neural network architecture was used for training, and iteration curves of different image enhancement methods were obtained, as shown in Fig. 7. As can be seen from Fig. 7, the balance was reached after about 150,000 iterations using different enhancement methods. The different enhancement algorithms used here are multi-scale Retinex with color restoration (MSRCR) (Wang et al., 2017a), multi-scale Retinex with chromaticity preservation (MSRCP) (Guo et al., 2021), and RetinexNet (Tang et al., 2022). MSRCR is an algorithm to improve MSR, that is, to add the function of color recovery on the basis of MSR, the expression is shown in formula Eq. (11). (11)R M S R C R x , y , σ = β log α I t x , y ∑ l = 1 3 I l x , y R M S R x , y , σ where R_MSRCR(x, y, σ) represents the output of the MSRCR process, which combines multi-scale Retinex (MSR) with a color restoration function.

β and α represent the scaling factors used for tuning the brightness and contrast of the image. I_s(x, y) represents the intensity of the image at position (x, y) for the s^th color channel. ∑ l = 1 3 I l x , y represents the sum of intensities across all three color channels (typically Red, Green, and Blue). R_MSR(x, y, σ) represents the multi-scale Retinex output, which enhances the contrast and dynamic range of the image.

To address the color bias issue arising from MSRCR, the MSRCP algorithm processes the MSR on the intensity data of the image. Subsequently, it maps this data back to each channel based on the original RGB scale. In RetinexNet, the algorithm aims to mitigate or eliminate the influence of the incident light on the image by preserving the essential reflective attributes of the object. The theoretical enhancement algorithm’s application formula is depicted in Eq. (12). (12)S x , y = R x , y ∗ L x , y .

S(x,y) represents the final enhanced image output at pixel location (x,y). R(x,y) represents the reflectance component of the image, which is independent of illumination and represents the intrinsic properties of the objects in the scene. L(x,y) represents the illumination component, which represents the varying light conditions in the scene.

Results

As shown in Fig. 7, the iteration curve of the original image fluctuates greatly in the early stage, while the iteration curve based on other image enhancement algorithms is relatively smooth. In addition, the trend of HE is close to that of MSRCR and MSRCP. However, careful observation of the development trend in the early and late stages shows that the enhancement method of HE has an obvious enhancement process, and the overall recognition success rate is slightly higher than that of other image enhancement methods. Therefore, the HE enhancement method adopted in this paper has played a good role.

The average recognition success rates after the algorithm stabilizes are depicted in Fig. 8. As shown, compared to other image enhancement methods, HE achieves the highest recognition success rate, with a 14.4% increase in accuracy compared to the original image. Additionally, the recognition success rates of image enhancement algorithms using MSRCR and MSRCP are slightly lower than HE, demonstrating their effectiveness to some extent.

While the Retinex method yields a recognition success rate higher than that of the original image, the improvement is minimal and notably lower overall compared to the other three methods. Therefore, based on the final average recognition rate results, the enhanced HE exhibits robust recognition performance compared to various image enhancement algorithms.

In addition, we conducted experiments on a feature extraction method based on CNN and Transformer dual-flow models using two datasets. We compared our method with several other prominent feature models including CNN (Munadi et al., 2020), GoogleNet (Chen et al., 2023), convolutional neural network and support vector machine (CNN + SVM) (Khairandish et al., 2022), ResNet (Wu, Shen & Hengel, 2019), and Transformer (Zhao et al., 2021), and evaluated their performance. As depicted in Fig. 9, our method achieves the highest values across all evaluation metrics compared to other algorithms.

The significant improvement in P, R and F1 score (F) values by more than 5% with our method is primarily attributed to enhancing the functionalities of CNN and Transformer, integrating them for feature extraction, and introducing an attention mechanism module to enhance feature extraction effectiveness. Consequently, our algorithm surpasses separate CNN and Transformer models by more than 10% and 5% in P, R, and F values, respectively. Compared to GoogleNet and ResNet, our approach utilizes a residual dense module at the core of the CNN branch to achieve multi-scale local feature extraction, resulting in notable improvements in P, R, and F values as observed in experimental results.

Although adding SVM improves accuracy in the CNN + SVM method compared to a standalone CNN model, our proposed algorithm consistently demonstrates stable improvements in relative accuracy across the two datasets post-experimentation.

To validate the effectiveness of the CTGAN algorithm, we conducted a comprehensive analysis by evaluating the PSNR and SSIM metrics across various algorithms using the DesignNet and AVA datasets. These metrics are critical indicators of image quality, where higher values denote superior performance. Table 3 presents the evaluation results, with the best performance highlighted in bold and the top performance underlined.

Quantitative comparisons reveal CTGAN’s superiority in both datasets. On the DesignNet dataset, CTGAN achieved a PSNR of 25.85, surpassing ResNet (25.19) and Transformer (24.24). The improvement of 0.66 PSNR over ResNet indicates CTGAN’s superior ability to reduce noise and enhance image details. On the AVA dataset, CTGAN’s PSNR of 26.86 is 1.65 higher than ResNet and Transformer, highlighting its effectiveness in maintaining high-fidelity images even on diverse datasets. For SSIM, CTGAN achieved 0.88 on the DesignNet dataset, compared to ResNet’s 0.83 and Transformer’s 0.85, reflecting CTGAN’s capability in preserving structural information and perceptual quality. On the AVA dataset, CTGAN’s SSIM of 0.90 outperforms both ResNet (0.87) and Transformer (0.88), reinforcing its strength in capturing fine details and textures.

Qualitative comparisons further underscore CTGAN’s advantages. Although CNN+SVM shows improved results over CNN alone, with a notable increase in PSNR and SSIM, CTGAN significantly surpasses these enhancements. The hybrid model of CNN+SVM achieves a PSNR of 23.28 and 23.50 on DesignNet and AVA, respectively, but falls short when compared to CTGAN. This indicates that while SVM aids in improving image quality, the generative approach of CTGAN offers a more substantial enhancement by generating finer details and reducing artifacts. GoogleNet’s performance is the lowest among the compared models, with a PSNR of 18.27 and 17.85 and SSIM of 0.66 and 0.67 on DesignNet and AVA datasets, respectively. The traditional network architecture without residual connections limits its ability to compete with more advanced models like ResNet and CTGAN, highlighting the need for residual structures and generative adversarial frameworks for superior image quality.

ResNet and Transformer networks perform commendably, with ResNet showing strong PSNR and SSIM, and Transformer slightly better in SSIM. However, CTGAN surpasses both in all metrics, which can be attributed to the fusion of CNN and Transformer networks, coupled with the generative adversarial approach that enhances feature representation and image quality. The significant performance improvement of CTGAN demonstrates the efficacy of integrating GANs with advanced network architectures like CNN and Transformer. The generative component of CTGAN allows for better feature extraction and noise reduction, leading to higher fidelity in the reconstructed images. Moreover, the adversarial training helps in refining details that are crucial for both quantitative metrics and visual quality.

The ablation study results are illustrated in Table 4.

Original image (Baseline): The image without any processing.

HE only: Image processed with histogram equalization only.

CNN flow: Image enhanced using only CNN feature extraction.

Transformer flow: Image enhanced using only Transformer feature extraction.

CNN + Transformer: Combination of CNN and Transformer for feature extraction and enhancement.

CTGAN (Full Model): Full CTGAN model, including HE processing, CNN flow, Transformer flow, and GAN generator.

The results of the ablation study provide critical insights into the efficacy of various image enhancement strategies. The baseline image, which is left unprocessed, serves as a control, offering a moderate level of quality, evidenced by a PSNR of 20.45 dB and an SSIM of 0.650. This establishes a fundamental benchmark against which enhancement techniques are evaluated. The use of HE alone substantially improves image quality, as reflected by notable increases in both PSNR and SSIM scores. These improvements suggest that HE effectively enhances contrast and detail visibility, though it falls short of achieving the superior metrics delivered by more sophisticated methods.

When CNNs are incorporated into the enhancement process, a further elevation in image quality is observed. The increase in PSNR and SSIM from the HE-only method to the CNN-enhanced approach implies that CNNs contribute crucial features that improve detail and clarity. This aligns with the well-established role of CNNs in identifying and amplifying intricate patterns within images. Similarly, the application of Transformer-based feature extraction results in comparable improvements to those achieved with CNNs, albeit with slight variations in performance metrics. This suggests that Transformers, known for their capacity to capture long-range dependencies and contextual information, also contribute positively to image quality. However, the performance of the Transformer-only approach does not exceed that of the CNN-based method.

The combination of CNN and Transformer techniques yields the most significant improvements, demonstrating that the integration of these methods harnesses their respective strengths in feature extraction. This joint approach capitalises on the CNN’s ability to extract detailed features and the Transformer’s capability for contextual understanding, leading to the highest observed PSNR and SSIM values. Ultimately, the full CTGAN model, which integrates histogram equalization, CNN and Transformer feature extraction, and GAN-based generation, achieves the best performance across all metrics. This comprehensive strategy underscores the synergetic effects of combining multiple advanced techniques, indicating that the CTGAN model not only enhances image quality through individual stages but also refines the final output via generative adversarial methods.