A HYBRID MULTISCALE CONVOLUTION NEURAL NETWORK WITH ATTENTION AND TEXTURE FEATURES FOR IMPROVED IMAGE CLASSIFICATION Irpan Adiputra Pardosi, Tengku Henny Febriana Harumy, Syahril Efendi Eastern European Journal of Enterprise Technologies, 2025 The object of this study is the classification of low-resolution and multi-class images, represented by the CIFAR-10 benchmark dataset. It is challenging to accurately classify low-resolution and multi-class images because traditional CNNs usually have trouble identifying both global and complex texture patterns. To address this issue, this study employs the CIFAR-10 dataset as a representative benchmark for real-world scenarios where image quality is limited, such as in low-cost medical imaging, remote sensing, and security surveillance systems. The limited discriminability of traditional CNNs in these situations is the primary issue addressed. The proposed method employs three parallel convolutional streams with distinct kernel sizes (3 × 3, 5 × 5, and 7 × 7) to capture hierarchical spatial patterns, followed by the integration of two attention mechanisms – squeeze-and-excitation and convolutional block attention module – that adaptively emphasize the most relevant spatial and channel-wise information. In addition, structural texture descriptors such as Gray-level co-occurrence matrix, local binary pattern, and Gabor filters are computed independently and later fused with the deep representations to enrich the feature space. Experiments were carried out on the CIFAR-10 dataset under varying levels of class complexity: 10, 5, and 3 categories. The results reveal that the hybrid approach significantly improves precision, recall, and F1-score across all scenarios, with the highest accuracy of 90.87% obtained when only three classes are involved. These improvements are explained by the complementary nature of deep and handcrafted features, which together enable the model to learn both global semantics and fine-grained local textures can achieve higher classification accuracy, improved reliability, and reduced misclassification errors, ultimately enhancing the effectiveness of applications ranging from medical decision support to intelligent surveillance.
BIG DATA ANALYTICS FOR SEASONAL CROP PATTERNS: INTEGRATING MACHINE LEARNING TECHNIQUES Roni Yunis, Arwin Halim, Irpan Adiputra Pardosi Eastern European Journal of Enterprise Technologies, 2024 This study addresses the challenge of predicting rice growing season lengths, crucial for agricultural planning in tropical regions. Climate variability and season timing create uncertainties in decision-making, and while machine learning is widely used in agriculture, a gap persists in integrating spatial-temporal data for accurate season length prediction and region-specific pattern analysis influenced by rainfall. Using a combination of Random Forest algorithms with hyperparameter optimization (grid search), and clustering techniques such as PCA, K-Means, and Hierarchical Clustering, this study analyzes key features such as the start of the season (SOS), end of the season (EOS), and their significance indicators (sig_sos and sig_eos). The findings reveal a strong correlation (0.98) between SOS and EOS, with an optimal growing season ranging from day 93 to day 207 (113.82 days). The Random Forest model, optimized with Grid Search, achieved a MSE of 28.9474 and an R2 of 0.8636, showing an outstanding predictive result. SHAP and LIME analyses identified sos and eos as the most influential predictors, while cluster analysis highlighted three distinct growing season groups characterized by variations in rainfall and seasonal stability. These results underscore the importance of understanding localized agricultural conditions and provide actionable insights for optimizing planting schedules, resource allocation, and climate adaptation strategies. By integrating advanced machine learning techniques with spatial-temporal data, this study establishes a foundation for improving agricultural resilience and sustainability in the face of climate variability
Implementation of Discrete Cosine Transform and Permutation-Substitution Scheme Based on Henon Chaotic Map for Images Andrean Lius, Eric, Irpan Adiputra Pardosi, Hernawati Gohzali 2022 7th International Conference on Informatics and Computing Icic 2022, 2022 The use of images from the internet that are not obtained from their creators is a copyright infringement. Illegal access to confidential images can be detrimental to the owner of the image. Things like this can be prevented by securing digital images through steganography and encryption processes. The methods that can be used to perform steganography and secure encryption of images are the Discrete Cosine Transform (DCT) method and A New Permutation-Substitution Scheme Based on Henon Chaotic Map for Image Encryption. Encryption uses the Henon Map algorithm to generate a key and the encryption process is carried out by randomizing each pixel in the image then XOR is performed to encrypt the color of the image. The DCT method is a lossy compression scheme where $\\mathrm{N}\\times\\mathrm{N}$ blocks are transformed from the spatial domain to the DCT domain. The results of this study can hide and secure the image using a Discrete Cosine Transform and A New Permutation-Substitution Scheme Based on Henon Chaotic Map. The best combination of key-value $\\mathrm{X}_{0}=0.017$ and $\\mathrm{Y}_{0}=0.061$. The result stego image is resistant to 0.1% noise attacks but is not resistant to 1% noise attacks and changes in contrast because the value of pixel changes too much so it can affect the message image that hides in the cover image.
Grayscale Image Quality Analysis Result of Noises Reduction using Adaptive Fuzzy Filter (AFF) and Spatial Median Filter (SMF) Against Image Depth Variations Irpan Adiputra Pardosi, Ali Akbar Lubis Journal of Physics Conference Series, 2019 Removing salt and pepper noise in the image will have an impact on information in the image, with a greater percentage of noise, the changes in the results image will also be large, this is very likely different from the types of grayscale images with different depths such as 8, 16 and 24 bit. The percentage of noise that will be used starts from 45%, this is based on previous studies which reveal the ability of noise reduction algorithms to be able to work a maximum of below 20%, so that it needs to be studied more deeply the performance of the algorithm and its impact on images with greater noise. From the results of previous research studies, the Spatial Median Filter algorithm and Adaptive Fuzzy Filter algorithm can reduce noise with the maximum percentage of noise below 45% in the 8 Bit image, but leave some noise after being reduced. The test results on grayscale image obtained the average overall results quality Adaptive Fuzzy Filter algorithm is better than Spatial Median Filter in terms of image quality using PSNR but the diversity of information is even more reduced compared to the SMF algorithm referring to the entropy shannon value.
Noise Reduction for Big Percentage Salt and Pepper Using a Combination of Algorithm in 24-Bit Color Image Irpan Adiputra Pardosi, Ali Akbar Lubis, Hernawati Gohzali Proceedings of 2019 4th International Conference on Informatics and Computing Icic 2019, 2019 Problem of salt noise in images can be reduced partially or in whole with various algorithms but will have an impact on the diversity of information in the image such as the color combination between RGB and image quality. With a large percentage of noise, the changes in the resulting image will also be large, especially in the 24bit image. This study specifically addresses the type of salt and pepper noise with a percentage of noise greater than 45% in bitmap color images, based on previous research which revealed the ability of noise reduction algorithms with Adaptive Fuzzy Filter (AFF) better than Spatial Median Filter (SMF) in terms of image quality with the best Peak Signal to Noise Ratio (PSNR) 28.10 dB in 24bit color images but left some noise after being reduced. This study will examine the stages of the noise reduction process with the aim of improving image quality up to close to 40 dB, starting from a combination of both algorithms (AFF and SMF) and the process of reducing noise with the same algorithm 2 times. Test results on 24bit bitmap color images for salt noise percentage 45%, obtained results image quality with the AFF algorithm is better than the SMF algorithm and the combination of both algorithms or the process of repeating the reduction with the same algorithm 2 times. The AFF algorithm is not able to reduce noise reduction to zero but the image quality is close to 40 dB which is 30.57 dB.
Combination of Steganography with K Means Clustering and 256 AES Cryptography for Secret Message Ali Akbar Lubis, Ronsen Purba, Irpan Adiputra Pardosi Proceedings of 2019 4th International Conference on Informatics and Computing Icic 2019, 2019 Steganography is a technique that evokes secret messages on the media. The cover requested for the message does not arouse suspicion, while cryptography is a senior art and science that maintains the confidentiality of data. Least Significant Bit (LSB) is one of the steganography methods for hiding secret messages carried out at the last bit in several bytes of media dataset values. The location of LSB embedding in the image cannot be randomized. This makes the location of the message that will be inserted (embedding) easier to detect which results in a decrease in the quality of the results of steganography and the similarity of the data inserted. To overcome the weaknesses of the LSB steganography method, improve the K Means Clustering method on steganography and AES 256 for message security. Grouping is a method that is carried out to select the object area, then the embedding of the message is done on the object obtained. As additional security, cryptography used is AES 256.
Improvement of Degraded Object Shapes Based on Skeletonizing and Recognizing Using Geometry Pahala Sirait, Irpan Adiputra Pardosi, Alex Ciawi, Hendra Tandiono, Kelvin Proceedings of 2019 4th International Conference on Informatics and Computing Icic 2019, 2019 Remote sensing technology normally uses for the provision of regional information to support decision making in area utilization, regional boundary determination, infrastructure mapping, etc. Through satellite imagery where it very sensitive to noise causing a shape building objects are degraded and cannot be recognized clearly caused by other objects covering observed building objects and there are holes due to pixel identification errors due to color differences so that the shape of the building object produced is inaccurate and only visible objects in certain parts. Methodology of this research is divided into two stages, the first stage of improvement of degraded building object forms, including separation of buildings with other objects using the K-Means Clustering algorithm (k = 2), Filling Region to cover holes, and Skeletonizing morphology to produce the framework used to search for endpoints that are lost when the indentation leads in, so that in the end the segmentation of the building object line will be formed perfectly through these endpoints; the second stage of the introduction of the shape of the building object using the Harris Corner Detector method to recognize shapes based on the total vertex. The results showed that using Skeletonizing, degraded building objects were able to be refined and can be identified based on angles with relatively high accuracy in 71%.
