Frequent Pattern Mining approach to Image Compression

The paper focuses on Image Compression, explaining efficient approaches based on Frequent Pattern Mining(FPM). The proposed compression mechanism is based on clustering similar pixels in the image and thus using cluster identifiers in image compression. Redundant data in the image is effectively handled by replacing the DCT phase of conventional JPEG through a mixture of k-means Clustering and Closed Frequent Sequence Mining. To optimize the cardinality of pattern(s) in encoding, efficient pruning techniques have been used through the refinement of Conventional Generalized Sequential Pattern Mining(GSP) algorithm. We have proposed a mechanism for finding the frequency of a sequence which will yield significant reduction in the code table size. The algorithm is tested by compressing benchmark datasets yielding an improvement of 45% in compression ratios, often outperforming the existing alternatives. PSNR and SSIM, which are the image quality metrics, have been tested which show a negligible loss in visual quality.

💡 Research Summary

The paper proposes a novel image compression framework that replaces the traditional Discrete Cosine Transform (DCT) and quantization stages of JPEG with techniques drawn from data mining, specifically Frequent Pattern Mining (FPM). The method operates entirely in the spatial domain and consists of four main phases: (1) Color‑channel clustering – the input RGB image is split into its three channels, and each channel is clustered using k‑means into k groups. Each pixel is replaced by the identifier of its cluster, thereby reducing the per‑pixel bit requirement from the original 8 bits to log₂k bits (or fewer if a compact identifier scheme is used). The cluster centroids serve as representative pixel values for later reconstruction. (2) Sequence database construction – the matrix of cluster identifiers is read row‑wise, treating each row as a sequence of symbols. These sequences constitute a database D for mining. (3) Closed Frequent Sequence Mining (CFSM) – a modified Generalized Sequential Pattern (GSP) algorithm is employed. Unlike the classic GSP, which first extracts all frequent sequences and then prunes non‑closed ones, the authors integrate a pruning step at each level to keep only closed frequent sequences. Moreover, they introduce a “modified support” (s_mod) concept to handle overlapping occurrences: the longest frequent sequences are matched first, removed from the database, and the supports of remaining shorter sequences are recomputed. This non‑overlapping counting dramatically reduces the size of the code table. (4) Huffman coding – the set of closed frequent sequences together with their modified supports is encoded using a Huffman tree. Because longer sequences typically have higher support, they receive shorter codewords, further compressing the data. The final compressed bitstream consists of Huffman codes that replace the original identifier matrix.

Decompression simply reverses these steps: Huffman decoding restores the identifier matrix, which is then mapped back to the cluster centroids to reconstruct the RGB image.

The authors evaluate the approach on standard 512 × 512 benchmark images (Lena, Peppers, Baboon, Boat) using a Python implementation on an Intel i5‑4210 CPU with 4 GB RAM. They vary the number of clusters k and the minimum support threshold α to study their impact. Results show that with k = 5 and α = 3, the Lena image is reduced from 512 bits (for an 8 × 8 × 8 block) to 129 bits, achieving a compression ratio of 3.968×. Across the test set, compression ratios improve by roughly 30 %–45 % compared with conventional JPEG, while Peak Signal‑to‑Noise Ratio (PSNR) remains above 30 dB and Structural Similarity Index (SSIM) exceeds 0.95, indicating negligible visual degradation.

The paper highlights several strengths: (i) a clear integration of clustering and pattern mining that directly exploits spatial redundancy; (ii) the modified support mechanism that eliminates double‑counting of overlapping patterns, thereby shrinking the code table; (iii) the use of Huffman coding to adaptively assign short codes to the most frequent, longest patterns. However, it also acknowledges limitations. The combined k‑means and CFSM steps are computationally intensive, especially for high‑resolution images, and the per‑channel independent processing ignores inter‑channel color correlations, potentially causing subtle color artifacts. Moreover, the compression performance is sensitive to the choice of k and α, requiring careful tuning or adaptive selection mechanisms.

Future work suggested includes: (a) accelerating clustering with mini‑batch or GPU‑based k‑means; (b) extending the mining stage to jointly consider all three color channels, possibly via multi‑dimensional sequential pattern mining; (c) incorporating hardware acceleration for the mining and coding phases; and (d) developing automatic parameter selection strategies based on image statistics. Overall, the study demonstrates that data‑mining concepts, when thoughtfully adapted, can serve as a viable alternative to traditional transform‑based image compression, offering competitive compression ratios with minimal quality loss.