Unsupervised autoencoder detection with a local context-based correction for mixed pixels in thermal images

Thermal imaging is important in areas like precision agriculture, surveillance, and remote sensing because it gives temperature-based information for monitoring environmental conditions. However, mixed pixels often reduce the accuracy of analysis in many of these applications. A mixed pixel is a fundamental distortion where a single pixel captures the blended thermal signature of several different surfaces. Unlike the most common distortions, such as noise or blur, mixed pixels distort the thermal values, leading to significant analytical inaccuracies. This work proposes a novel, label-free unsupervised framework that does not require annotated mixed-pixel labels to address this challenge and it works in two distinct phases. The first step is to use a convolutional autoencoder, which finds the mixed pixels by determining regions with a higher reconstruction error. The threshold was first empirically calibrated to guarantee zero false positives on a verified single-object control dataset of thermal images and then finally applied to the complex multi-object agricultural scenarios to ensure boundary focused mixed pixel detection. In the second step, a local context-based refinement algorithm corrects each flagged mixed pixel by analyzing the immediate neighborhood. The process involves finding the number of contributing thermal profiles using a color similarity metric and then reassigning the pixel value to that of the prominent profile, which is identified by K-means clustering. We validate the framework through a comprehensive evaluation where corrected images gave significant enhancement in structural similarity. The SSIM and MSE scores are computed between the original images and the autoencoder reconstruction of the corrected thermal images to quantify structural consistency and pixel-level stabilization, rather than against an external ground-truth image. The SSIM value increased from 0.4184 to 0.8201, and the MSE value decreased from 0.1088 to 0.0022 on Dataset III, while similar trends can be observed in other datasets. Besides, a CNN-based classification study on correct image datasets showed accuracy improvement of 7–12%, proving that the smoothing of mixed pixels improves the model performance. These results point towards the fact that addressing the mixed-pixel distortion significantly raises the reliability and interpretability of thermal Imaging for various real-world applications.