This study first proposes a variable template matching algorithm for anomaly detection of images. Variational forms of template for defect with multiple scales, rotations and perspective transformations are included to improve its variational robustness. The normalized cross correlation between the template and the sliding window on the image is computed as the matching result. Secondly, it proposes a kernel density algorithm, in which a lower kernel density index (intensity percentile/range) of the sliding window indicates a potential anomaly. Lastly, it proposes a gradient convolution algorithm. These three traditional computer vision algorithms are implemented for anomaly detection of a group of biological images, and results are compared with that of the convolutional neural network ResNet-50. Results show that the variable template matching algorithm achieves superior performance (true positive 82% and false positive 4.9%) than both the kernel density and gradient convolution algorithms. Its detection rate is lower than the prediction of ResNet-50 (true positive 86%), but it is much faster and does not need any train images as an unsupervised learning. Therefore, it can be a potential candidate for object detection of small dataset or for a quick solution.