The invention discloses an improved particle swarm algorithm based on support vector machine regression. The improved particle swarm algorithm includes: randomly taking a point near the current globally optimal particle position, predicting an optimal random point by using support vector machine regression, then replacing a non-optimal particle, and carrying out next iteration. Along with continuous increase of input regression information, prediction becomes more and more accurate. When a bird flock searches for food, the bird flock always moves towards a known optimal position, the bird flock has a certain probability and does not move towards the optimal position, an intelligent person exists in the bird flock, and the position of a better point near the optimal position can be pre-judged according to the point where the bird flock passes through. A random value range of an original particle swarm algorithm is changed from [0, 1] to [-1, 1], a value is randomly taken near an optimalvalue after an optimal particle position is calculated each time, and a support vector machine is used for regression prediction of the optimal value, so that the globality and locality of a group are enhanced, and the global and local optimization capability can be effectively enhanced. The improved particle swarm algorithm can be specifically applied to complex optimization problems such as function optimization, planning problems, mode recognition and image processing problems.