Fruits target recognition is one of the most important steps to realize agricultural automation. In the process of fruit recognition, because of the influence of overlap and occlusion, the target recognition is difficult, and the rate of recognition is not high. The paper uses spectral clustering algorithm to solve the problem of overlapped fruit in natural environment. Then the identification and location of fruit are realized by randomized hough transform. In view of the large number of computation and the slow operation speed of the traditional algorithm, this paper proposes an improved spectral clustering algorithm based on Mean Shift and sparse matrix principle. Firstly, the image is pre-segmented using the mean shift algorithm. Mean shift is a non-parametric estimation method for density gradient. The algorithm is essentially an iteration. Calculate the offset, move the point according to the offset, and repeat the above steps until the offset is zero. Most of the background pixels are removed by mean shift algorithm, and the removing is prepared for reducing the computational complexity of the spectral clustering algorithm. And then the useful information is extracted, which is the description of the similarity between the pairs of pixels in the image, and the extracted image feature information is mapped into a sparse matrix. The K-means algorithm is used to classify it into classes, and the final classification result is obtained to realize the re-segmentation of the reprocessed image. Then the color of the image segmentation area is restored, the edge contour is extracted by using a color vector gradient and the randomized hough transform is used on the resulting contour image, and the radius parameter range during the detection process is set to further accelerate the speed of the algorithm. The center coordinates and radius of the target can be obtained through the detection. Thereby the overlapped green apples are recognized. Finally, the algorithm has the high coincidence degree of 9541%, the low error rate of 459% and the false detection rate of 305% through experimental analysis and algorithm comparison, and the algorithm meets the practical application requirements.